def draw(self):
if self.markers is not None:
self.result[:] = self.markers
cv2.watershed(self.img, self.result)
img = MARKER_COLORS[self.result]
self.draw_image(img, self.result_artist, draw=False)
self.draw_image(self.img)
def segment_on_dt(a, img, gray):
border = cv2.dilate(img, None, iterations=5)
border = border - cv2.erode(border, None)
dt = cv2.distanceTransform(img,cv2.DIST_L2,5)
plt.subplot(3,3,4)
plt.imshow(dt),plt.title('dt'),plt.xticks([]), plt.yticks([])
dt = ((dt - dt.min()) / (dt.max() - dt.min()) * 255).astype(np.uint8)
_, dt2 = cv2.threshold(dt, 0, 255, cv2.THRESH_BINARY)
dt2 = cv2.erode(dt2, None, iterations=2)
# dt2 = cv2.adaptiveThreshold(dt, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 5, 2)
# dt1 = peak_local_max(gray, indices=False, min_distance=10, labels=img, threshold_abs=5)
# dt2 = peak_local_max(dt, indices=False, min_distance=5, labels=img, threshold_abs=0)
lbl, ncc = label(dt2)
plt.subplot(3,3,5)
plt.imshow(dt2),plt.title('localMax'),plt.xticks([]), plt.yticks([])
# plt.subplot(3,3,6)
# plt.imshow(ncc),plt.title('ncc'),plt.xticks([]), plt.yticks([])
lbl = lbl * (255/ncc)
# Completing the markers now.
lbl[border == 255] = 255
lbl = lbl.astype(np.int32)
cv2.watershed(a, lbl)
lbl[lbl == -1] = 0
lbl = lbl.astype(np.uint8)
plt.subplot(3,3,6)
plt.imshow(lbl),plt.title('lbl_out'),plt.xticks([]), plt.yticks([])
return 255 - lbl
def watershed(self,image, marker):
"""
Applies opencv's watershed method iteratively to an input image. An initial marker containing
preliminary information on which pixels are foreground serves as additional input.
The initial marker can be based on user input (color-picking), or can be constructed
with an automatic marker strategy. The marker decides from which pixels the flooding
in the watershed method may start. Finally, the marker is used to obtain a mask
classifying every pixel into foreground or background.
Args:
image: An input image which is not altered
marker: A marer suitable for use with opencv's grabcut
iterations: The number of iterations grabcut may update the marker
Returns:
A mask image classifying every pixel into foreground or background
"""
if len(marker.shape) == 3:
marker = marker[:,:,0]
# convert the marker to the format watershed expects
marker = np.int32(marker)
# Use watershed to decide how to label the yet undecided regions.
# This may produce may foreground areas labeld with different integers.
cv2.watershed(image, marker)
# Obtain the final mask by thresholding. Different foreground regions have different positive values.
# We are only intrested in global foreground so we set all of them to be white, i.e. 255.
mask_watershed = cv2.convertScaleAbs(marker)
return mask_watershed
def watershed(image, grayed, edges, min_ratio, max_count):
""" Applies watershed algorithm to 'image' with markers derived
from 'edges'
Args:
image: original image
grayed: grayed and optionally blurred version of 'image'
edges: a binary image
min_ratio: only contours in 'edges' with an area bigger are used as
markers
max_count: maximum number of segments to derive
Returns segments, markers, count
"""
markers = edges.copy()
_, markers1, _ = extract_segments(
grayed,
markers,
min_ratio=min_ratio,
max_count=max_count
)
markers32 = numpy.int32(markers1)
cv2.watershed(image, markers32)
watersheded = cv2.convertScaleAbs(markers32)
_, edges = cv2.threshold(
watersheded,
1,
255,
cv2.THRESH_BINARY_INV
)
segments, markers, count = extract_segments(
grayed,
edges
)
return segments, markers, count
def subtract_back(self,frm):
#dst=self.__back__-self.__foreground__
temp=np.zeros((config.height,config.width),np.uint8)
self.__foreground__=cv2.blur(self.__foreground__,(3,3))
dst=cv2.absdiff(self.__back__,self.__foreground__)
#dst=cv2.adaptiveThreshold(dst,255,cv.CV_THRESH_BINARY,cv.CV_ADAPTIVE_THRESH_GAUSSIAN_C,5,10)
val,dst=cv2.threshold(dst,0,255,cv.CV_THRESH_BINARY+cv.CV_THRESH_OTSU)
fg=cv2.erode(dst,None,iterations=1)
bg=cv2.dilate(dst,None,iterations=4)
_,bg=cv2.threshold(bg,1,128,1)
mark=cv2.add(fg,bg)
mark32=np.int32(mark)
#dst.copy(temp)
#seq=cv.FindContours(cv.fromarray(dst),self.mem,cv.CV_RETR_EXTERNAL,cv.CV_CHAIN_APPROX_SIMPLE)
#cntr,h=cv2.findContours(dst,cv.CV_RETR_EXTERNAL,cv.CV_CHAIN_APPROX_SIMPLE)
#print cntr,h
#cv.DrawContours(cv.fromarray(temp),seq,(255,255,255),(255,255,255),1,cv.CV_FILLED)
cv2.watershed(frm, mark32)
self.final_mask=cv2.convertScaleAbs(mark32)
def segment_on_dt(a, img):
'''
find foreground
'''
fg=cv2.dilate(img,None, iterations=5)
fg=fg-cv2.erode(fg, None)
dt=cv2.distanceTransform(img, 2, 3)
dt=((dt-dt.min())/(dt.max()-dt.min())*255).astype(np.uint8)
_, dt=cv2.threshold(dt, 0, 255, cv2.THRESH_BINARY)
lbl, ncc=label(dt)
lbl=lbl*(255/ncc)
'''
Completing the markers now
'''
lbl[fg==255]=255
lbl=lbl.astype(np.int32)
cv2.watershed(a, lbl)
lbl[lbl==-1]=0
lbl=lbl.astype(np.uint8)
return 255-lbl
def segment_watershed(img):
gray=cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
ret, thresh=cv2.threshold(gray, 0,255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)
'''
foreground region
'''
fg=cv2.erode(thresh,None, iterations=2)
'''
background region
'''
bg=cv2.dilate(thresh, None,iterations=3)
ret, bg=cv2.threshold(bg,1,128,1)
'''
add both fg and bg
'''
marker=cv2.add(fg, bg)
'''
convert into 32SC1
'''
marker32=np.int32(marker)
'''
apply watershed
'''
cv2.watershed(img, marker32)
m=cv2.convertScaleAbs(marker32)
'''
threshold it properly to get the mask and perform bitwise_and with the input image
'''
ret, thresh=cv2.threshold(m,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
result=cv2.bitwise_and(img, img,mask=thresh)
return result
def watershed(self):
m = self.markers.copy()
cv2.watershed(self.img, m)
overlay = self.colors[np.maximum(m, 0)]
cv2.imshow('over', overlay)
vis = cv2.addWeighted(self.img, 0.5, overlay, 0.5, 0.0, dtype=cv2.CV_8UC3)
cv2.imshow('watershed', vis)
def detect_shirt2(self):
self.hsv=cv2.cvtColor(self.norm_rgb,cv.CV_BGR2HSV)
self.hue,s,_=cv2.split(self.hsv)
_,self.dst=cv2.threshold(self.hue,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
self.fg=cv2.erode(self.dst,None,iterations=3)
self.bg=cv2.dilate(self.dst,None,iterations=1)
_,self.bg=cv2.threshold(self.bg,1,128,1)
mark=cv2.add(self.fg,self.bg)
mark32=np.int32(mark)
cv2.watershed(self.norm_rgb,mark32)
m=cv2.convertScaleAbs(mark32)
_,m=cv2.threshold(m,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
cntr,h=cv2.findContours(m,cv.CV_RETR_EXTERNAL,cv.CV_CHAIN_APPROX_SIMPLE)
print len(cntr)
#print cntr[0].shape
#cntr[1].dtype=np.float32
#ret=cv2.contourArea(np.array(cntr[1]))
#print ret
#cntr[0].dtype=np.uint8
cv2.drawContours(m,cntr,-1,(255,255,255),3)
cv2.imshow("mask_fg",self.fg)
cv2.imshow("mask_bg",self.bg)
cv2.imshow("mark",m)
def watershed(self):
m = self.markers.copy()
cv2.watershed(self.img,m)
self.img[m == -1] = [255,0,0]
overlay = self.colors[np.maximum(m, 0)]
vis = cv2.addWeighted(self.img, 0.5, overlay, 0.5, 0.0, dtype=cv2.CV_8UC3)
self.DisplayandSave('watershed', vis)
def process(self, args):
"""
Use the Watershed algorithm from the opencv package
to the selected color channels of the current image.
Args:
| *args* : a list containing image array and Graph object
"""
# convert image to grayscale and threshold it with the best threshold
# value usinf otsu binarization
image = args[0]
im_gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
ret, thresh = cv2.threshold(im_gray, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
# erode will be used to create the foreground region (marked with 255)
fg = cv2.erode(thresh, None, iterations=self.fgiter)
# dilate will be used to reduce the background region. Black region
# will be set to 128 and is marked as background
bgt = cv2.dilate(thresh, None, iterations=self.bgiter)
ret, bg = cv2.threshold(bgt, 1, 128, 1)
# create marker with foreground and background region
marker = cv2.add(fg, bg)
marker32 = np.int32(marker)
# use watershed and convert result back
cv2.watershed(image, marker32)
m = cv2.convertScaleAbs(marker32)
# threshold to get the mask
ret, thresh = cv2.threshold(m, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
self.result['img'] = cv2.bitwise_and(image, image, mask=thresh)
def get_fish_contour(data, cal):
color_image = cv2.cvtColor(data.array, cv2.COLOR_GRAY2BGR)
delta_image(data, cal)
threshold_by_type(data)
erode(data)
distance_transform(data)
threshold_by_type(data, otsu=False, thresh=5)
markers = data.array.copy()
dilate(data, iterations=7)
border = data.array.copy()
border = border - cv2.erode(border, None)
markers, num_blobs = ndimage.measurements.label(markers)
markers[border == 255] = 255
markers = markers.astype(np.int32)
cv2.watershed(image=color_image, markers=markers)
markers[markers == -1] = 0
markers = 255 - markers.astype(np.uint8)
image = FFImage(markers, meta=data.meta, log=data.log)
annotate_hu_moments(image)
image.meta['timestamp'] = time.time()
return image.meta
def segment_on_dt(img):
#http://stackoverflow.com/questions/11294859/how-to-define-the-markers-for-watershed-in-opencv
img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
_, img_bin = cv2.threshold(img_gray, 0, 255,cv2.THRESH_OTSU)
img_bin = cv2.morphologyEx(img_bin, cv2.MORPH_OPEN,numpy.ones((3, 3), dtype=int))
border = cv2.dilate(img_bin, None, iterations=5)
border = border - cv2.erode(border, None)
dt = cv2.distanceTransform(img_bin, 2, 3)
dt = ((dt - dt.min()) / (dt.max() - dt.min()) * 255).astype(numpy.uint8)
_, dt = cv2.threshold(dt, 180, 255, cv2.THRESH_BINARY)
lbl, ncc = label(dt)
lbl = lbl * (255/ncc)
# Completing the markers now.
lbl[border == 255] = 255
lbl = lbl.astype(numpy.int32)
cv2.watershed(img, lbl)
lbl[lbl == -1] = 0
lbl = lbl.astype(numpy.uint8)
result = 255 - lbl
result[result != 255] = 0
result = cv2.dilate(result, None)
img[result == 255] = (0, 0, 255)
return img
def remove_pectoral(self, img, breast_mask, high_int_threshold=.8,
morph_kn_size=3, n_morph_op=7, sm_kn_size=25):
'''Remove the pectoral muscle region from an input image
Args:
img (2D array): input image as a numpy 2D array.
breast_mask (2D array):
high_int_threshold ([int]): a global threshold for high intensity
regions such as the pectoral muscle. Default is 200.
morph_kn_size ([int]): kernel size for morphological operations
such as erosions and dilations. Default is 3.
n_morph_op ([int]): number of morphological operations. Default is 7.
sm_kn_size ([int]): kernel size for final smoothing (i.e. opening).
Default is 25.
Returns:
an output image with pectoral muscle region removed as a numpy
2D array.
Notes: this has not been tested on .dcm files yet. It may not work!!!
'''
# Enhance contrast and then thresholding.
img_equ = cv2.equalizeHist(img)
if high_int_threshold < 1.:
high_th = int(img.max()*high_int_threshold)
else:
high_th = int(high_int_threshold)
maxval = self.max_pix_val(img.dtype)
_, img_bin = cv2.threshold(img_equ, high_th,
maxval=maxval, type=cv2.THRESH_BINARY)
pect_marker_img = np.zeros(img_bin.shape, dtype=np.int32)
# Sure foreground (shall be pectoral).
pect_mask_init = self.select_largest_obj(img_bin, lab_val=maxval,
fill_holes=True,
smooth_boundary=False)
kernel_ = np.ones((morph_kn_size, morph_kn_size), dtype=np.uint8)
pect_mask_eroded = cv2.erode(pect_mask_init, kernel_,
iterations=n_morph_op)
pect_marker_img[pect_mask_eroded > 0] = 255
# Sure background - breast.
pect_mask_dilated = cv2.dilate(pect_mask_init, kernel_,
iterations=n_morph_op)
pect_marker_img[pect_mask_dilated == 0] = 128
# Sure background - pure background.
pect_marker_img[breast_mask == 0] = 64
# Watershed segmentation.
img_equ_3c = cv2.cvtColor(img_equ, cv2.COLOR_GRAY2BGR)
cv2.watershed(img_equ_3c, pect_marker_img)
img_equ_3c[pect_marker_img == -1] = (0, 0, 255)
# Extract only the breast and smooth.
breast_only_mask = pect_marker_img.copy()
breast_only_mask[breast_only_mask == -1] = 0
breast_only_mask = breast_only_mask.astype(np.uint8)
breast_only_mask[breast_only_mask != 128] = 0
breast_only_mask[breast_only_mask == 128] = 255
kernel_ = np.ones((sm_kn_size, sm_kn_size), dtype=np.uint8)
breast_only_mask = cv2.morphologyEx(breast_only_mask, cv2.MORPH_OPEN,
kernel_)
img_breast_only = cv2.bitwise_and(img_equ, breast_only_mask)
return (img_breast_only, img_equ_3c)
def split(self):
"""Split the region according to the normalized cut criterion.
Returns
-------
list of CutRegion
The regions created by splitting.
float
The normalized cut cost.
"""
if not cv2_available:
raise ImportError('OpenCV >= 2.4.8 required')
tmp_im = np.zeros(self.shape[0] * self.shape[1])
tmp_im[self.indices] = 1
labeled_array, num_features = ndimage.label(tmp_im.reshape(self.shape))
if num_features > 1:
labeled_array = labeled_array.reshape(-1)[self.indices]
segments = []
for i in range(1, num_features + 1):
idx = np.nonzero(labeled_array == i)[0]
segments.append(CutRegion(self.affinity_matrix[idx, :][:, idx],
self.indices[idx], self.shape))
return segments, 0.0
C = normcut_vectors(self.affinity_matrix, 1)
im = C[:, -2]
im -= im.min()
im /= im.max()
markers = -np.ones(self.shape[0] * self.shape[1]).astype('uint16')
markers[self.indices] = 0
markers[self.indices[im < 0.02]] = 1
markers[self.indices[im > 0.98]] = 2
markers = markers.reshape(self.shape)
vis2 = 0.5 * np.ones(self.shape[0] * self.shape[1])
vis2[self.indices] = im
vis2 *= (2 ** 8 - 1)
vis2 = cv2.cvtColor(np.uint8(vis2.reshape(self.shape)),
cv2.COLOR_GRAY2BGR)
markers = np.int32(markers)
cv2.watershed(vis2, markers)
cut = ndimage.morphology.binary_dilation(markers == 2).reshape(-1)[
self.indices]
cut[im > 0.98] = True
cost = self._normalized_cut_cost(cut)
for thresh in np.linspace(0, 1, 20)[1:-1]:
tmp_cut = im > thresh
tmp_cost = self._normalized_cut_cost(tmp_cut)
if tmp_cost < cost:
cost = tmp_cost
cut = tmp_cut
a = cut.nonzero()[0]
a = cut.nonzero()[0]
b = np.logical_not(cut).nonzero()[0]
return (
[CutRegion(self.affinity_matrix[x, :][:, x], self.indices[x],
self.shape) for x in [a, b] if len(x)],
self._normalized_cut_cost(cut)
)
def watershed(image, marker):
m = marker.copy()
cv2.watershed(image, m)
m[m != 1] = 0
m *= 255
points = cv2.findNonZero(m.astype(np.uint8))
bound_rect = cv2.boundingRect(points)
# x, y, w, h = bound_rect
return bound_rect
def watershed(img, marker, device, debug):
# Uses the watershed algorithm to detect boundry of objects
# Needs a marker file which specifies area which is object (white), background (grey), unknown area (black)
# img = image to perform watershed on needs to be 3D (i.e. np.shape = x,y,z not np.shape = x,y)
# marker = a 32-bit image file that specifies what areas are what (2D, np.shape = x,y)
cv2.watershed(img, marker)
device += 1
if debug:
print_image(marker, str(device) + 'watershed_img' + '.png')
return device, marker
def segment_watershed(image, window=None):
"""Segments an image using watershed technique.
Parameters
----------
image : (M, N, 3) array
Image to process.
window : tuple, (x, y, w, h)
Optional subwindow in image.
Returns
-------
(rects, display) : list, (M, N, 3) array
Region results and visualization image.
"""
import cv2
import numpy as np
if window:
subimage = np.array(image)
x, y, w, h = window
image = subimage[y:y + h, x:x + w]
rects, display = segment_edges(image, variance_threshold=100,
size_filter=0)
h, w = image.shape[:2]
initial = np.zeros((h, w), np.int32)
for i, rect in enumerate(rects):
cv2.drawContours(initial, [rect[4]], -1, 1, -1)
display = np.zeros(initial.shape, dtype=np.uint8)
segment_rects = []
for i, rect in enumerate(rects):
regions = np.zeros(initial.shape, dtype=np.uint8)
cv2.drawContours(regions, [rect[4]], -1, 2, -1)
regions = cv2.erode(regions, None)
regions = cv2.erode(regions, None)
markers = initial.copy()
markers[regions == 2] = 2
cv2.watershed(image, markers)
display[markers == 2] = 255
contours, hierarchy = _find_contours(regions.copy(),
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
segment_rects.extend([cv2.boundingRect(c) for c in contours])
if window:
new_rects = []
for rect in segment_rects:
dx, dy = 0, 0
new_rect = (rect[0] + x - dx, rect[1] + y - dy,
rect[2] + 2 * dx, rect[3] + 2 * dy)
new_rects.append(new_rect)
segment_rects = new_rects
return segment_rects, np.dstack(3 * [display])
def segment_on_dt(img):
dt = cv2.distanceTransform(img, 2, 3) # L2 norm, 3x3 mask
dt = ((dt - dt.min()) / (dt.max() - dt.min()) * 255).astype(numpy.uint8)
dt = cv2.threshold(dt, 100, 255, cv2.THRESH_BINARY)[1]
lbl, ncc = label(dt)
lbl[img == 0] = lbl.max() + 1
lbl = lbl.astype(numpy.int32)
cv2.watershed(cv2.cvtColor(img, cv2.COLOR_GRAY2BGR), lbl)
lbl[lbl == -1] = 0
return lbl
def get_keys_from_markers(img, markers_list):
#make the markers in the format required by the algorithm
h, w = img.shape[:2]
markers_array = np.zeros((h, w), np.int32)
boundary = 0
for i, mark in enumerate(markers_list):
markers_array[mark[1]][mark[0]] = i+1
boundary +=1
cv2.watershed(img, markers_array)
return markers_array, boundary
def remove_bg(self):
gray = cv2.cvtColor(self.img,cv2.COLOR_BGR2GRAY)
ret,thresh = cv2.threshold(gray,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
fg = cv2.erode(thresh,None,iterations = 2)
bgt = cv2.dilate(thresh,None,iterations = 3)
ret,bg = cv2.threshold(bgt,1,128,1)
marker = cv2.add(fg,bg)
marker32 = np.int32(marker)
cv2.watershed(self.img,marker32)
m = cv2.convertScaleAbs(marker32)
self.nobg_img = cv2.threshold(m,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)[1]
return self.nobg_img
def method2(self, image):
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
ret, thresh = cv2.threshold(gray,0,255,cv2.THRESH_BINARY + cv2.THRESH_OTSU)
fg = cv2.erode(thresh,None,iterations = 2)
bgt = cv2.dilate(thresh,None,iterations = 3)
ret,bg = cv2.threshold(bgt,1,128,1)
marker = cv2.add(fg,bg)
marker32 = np.int32(marker)
cv2.watershed(image,marker32)
m = cv2.convertScaleAbs(marker32)
ret,thresh = cv2.threshold(m,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
res = cv2.bitwise_and(image,image,mask = thresh)
return res
def find_optic_disc_watershed(img, P):
"""
Find optic disk in image using a watershed method.
:param img: BGR image
:param P: gray image
:return: optic_disc, Crs, markers, watershed
"""
assert img is not None and P is not None
#fore = cv2.cvtColor(P,cv2.COLOR_GRAY2BGR)
# create watershed
data_min, data_body_left, data_body, data_max_left = retina_markers_thresh(
P)
watershed = np.zeros_like(P).astype("int32")
mk_back, mk_body, mk_flare = 1, 2, 3
# background FIXMED use P.min() and aproaching to local maxima
watershed[P <= data_min] = mk_back
watershed[np.bitwise_and(P > data_body_left, P <
data_body)] = mk_body # main body
# find bright objects
flares_thresh = (P >= data_max_left).astype(np.uint8)
contours, hierarchy = findContours(
flares_thresh.copy(), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
mks_flare = []
for i, cnt in enumerate(contours):
index = mk_flare + i
mks_flare.append(index)
# Flares. this can be used approaching
watershed[contours2mask(cnt, shape=watershed.shape)] = index
# to local maxima, but brightest areas are almost
# always saturated so no need to use it
markers = watershed.copy()
# apply watershed to watershed
# FIXME perhaps the function should be cv2.floodFill?
cv2.watershed(img, watershed)
# different classification algorithms could be used using watershed
contours_flares = []
for mk_flare in mks_flare:
brightest = np.uint8(watershed == mk_flare)
contours, hierarchy = findContours(
brightest, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
contours_flares.extend(contours)
Crs = [(i, j) for i, j in [(convexityRatio(cnt), cnt)
for cnt in contours_flares] if i != 0] # convexity ratios
Crs.sort(reverse=True, key=lambda x: x[0])
candidate = Crs[-1]
ellipse = cv2.fitEllipse(candidate[1])
optic_disc = np.zeros_like(P)
cv2.ellipse(optic_disc, ellipse, 1, -1) # get elliptical ROI
return optic_disc, Crs, markers, watershed
def WATERSHED(Input_Image):
'''
Description: Computes watershed segmentation,based on mathematical morphology and flooding of regions from markers.
source: openCV
parameters: Input_Image : ndarray
Input image
marker: float
return: Segment_mask : ndarray (cols, rows)
Integer mask indicating segment labels.
'''
# read the input image
img = cv2.imread(Input_Image)
# convert to grayscale
g1 = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
# smooth the image
g = cv2.medianBlur(g1,5)
# Apply adaptive threshold
thresh1 = cv2.adaptiveThreshold(g,255,1,1,11,2)
thresh_color = cv2.cvtColor(thresh1,cv2.COLOR_GRAY2BGR)
# Apply dilation and erosion
bgt = cv2.dilate(thresh1,None,iterations = 3)
fg = cv2.erode(bgt,None,iterations = 2)
#thresholding on the background
ret,bg = cv2.threshold(bgt,1,128,1)
#adding results
marker = cv2.add(fg,bg)
#moving markers to 32 bit signed single channel
marker32 = np.int32(marker)
#segmenting
cv2.watershed(img,marker32)
m = cv2.convertScaleAbs(marker32)
ret,thresh = cv2.threshold(m,0,255,cv2.THRESH_BINARY+cv2.THRESH_OTSU)
res = cv2.bitwise_and(img,img,mask = thresh)
segments_watershed = cv2.cvtColor(res, cv2.COLOR_BGR2GRAY)
print("watershed number of segments: %d" % len(np.unique(segments_watershed)))
segments_watershed = segments_watershed.astype(np.int32)
#print ('segments_watershed datatype',segments_watershed.dtype )
return segments_watershed
def watershed(img, fg_marker):
fg_marker = cv2.dilate(fg_marker, None, iterations=5)
r, bg_marker = cv2.threshold(cv2.bitwise_not(cv2.dilate(fg_marker, None, iterations=10)), 30, 125, cv2.THRESH_BINARY)
marker = cv2.add(fg_marker, bg_marker, dtype=cv2.CV_32SC1)
cv2.imwrite('step7-markers.png', marker)
color = numpy.zeros((480, 640, 3), numpy.uint8)
color[:,:,0] = img
color[:,:,1] = img
color[:,:,2] = img
cv2.watershed(color, marker)
cv2.imwrite('step7-watersheds.png', marker)
return marker
def watershed_segmentation(bbox, image, rgb_proposal_mask, depth_proposal_mask, bbox_not_moved, watershed_last_frame_seed_mask):
# segmentation
watershed_mask_seed = create_watershed_seed(bbox, rgb_proposal_mask, bbox_not_moved, watershed_last_frame_seed_mask)
watershed_bg_mask = rgb_proposal_mask + depth_proposal_mask
# set the background to 1 the luggage pixel to the values found before
# the unknown pixel are still 0
#watershed_mask_seed = np.where(watershed_bg_mask == 0, 1, watershed_mask_seed)
watershed_mask_seed += 1
# apply watershed - result overwrite in mask
cv2.watershed(image, watershed_mask_seed)
# OUTPUT MASK FOR FURTHER STUDY
return np.where(watershed_mask_seed == 1, 0, 1)
def watershed(img, thresh):
# noise removal
kernel = np.ones((3,3), np.uint8)
opening = cv2.morphologyEx(thresh,cv2.MORPH_OPEN,kernel, iterations = 4)
# sure background area
sure_bg = cv2.dilate(opening,kernel,iterations=3)
#sure_bg = cv2.morphologyEx(sure_bg, cv2.MORPH_TOPHAT, kernel)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform,0.7*dist_transform.max(),255,0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg,sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers+1
# Now, mark the region of unknown with zero
markers[unknown==255] = 0
'''
imgray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
imgray = cv2.GaussianBlur(imgray, (5, 5), 0)
img = cv2.Canny(imgray,200,500)
'''
markers = cv2.watershed(img,markers)
img[markers == -1] = [255,0,0]
return sure_bg, sure_fg
def water(img, thresh):
kernel = np.ones((3,3),np.uint8)
opening = cv2.morphologyEx(thresh,cv2.MORPH_OPEN,kernel, iterations = 2)
# sure background area
sure_bg = cv2.dilate(opening,kernel,iterations=3)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening,2,5)
ret, sure_fg = cv2.threshold(dist_transform,0.7*dist_transform.max(),255,0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg,sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers += 1
# Now, mark the region of unknown with zero
markers[unknown==255] = 0
markers = cv2.watershed(img,markers)
img[markers == -1] = [255,0,0]
return sure_fg, sure_bg, markers, img
def Segmentation(image):
height, width = image.shape[0], image.shape[1]
temperatureImg = getTemperatureImg(image)
averTemperature = temperatureImg.sum() / (height * width)
data = cv2.calcHist([temperatureImg], [0], None, [256], [0, 255])
threshold = OSTU(data)
_, img = cv2.threshold(temperatureImg, 1.2 * threshold, 255, cv2.THRESH_BINARY)
fg = cv2.erode(img, None, iterations = 5)
_, img = cv2.threshold(temperatureImg, 0.8 * threshold, 255, cv2.THRESH_BINARY)
_, bg = cv2.threshold(cv2.dilate(img, None, iterations = 3), 1, 128, cv2.THRESH_BINARY_INV)
marker = cv2.add(fg, bg)
markers = np.int32(marker)
# in markers 255 is colorfulcolor, 128 is other, -1 is boundary
tpImg = np.rollaxis(np.array([temperatureImg] * 3), 0, 3)
cv2.watershed(np.array(tpImg, dtype='uint8'), markers)
return (markers, temperatureImg)
def watershed2(seem, img):
if (img.dtype != np.uint8):
img = img * 255
img = np.uint8(img)
ut.imgDes(img)
img = cv.medianBlur(img, 3)
rows, cols = img.shape[:2];
if np.size(img.shape) < 3:
img_a = cv.cvtColor(img, cv.COLOR_GRAY2RGB)
else:
img_a = img
print ('ddd')
ut.imgDes(img_a)
ret, markers = cv.connectedComponents(seem)
markers = markers + 1
markers[seem == 0] = 0
markers = cv.watershed(img_a, markers)
markers[img==0] = 0
return markers
cv2.circle(markedImg, pt, 9, (colorMark), -1)
cv2.circle(scene, pt, 9, colorPallet[colorMark], -1)
changedMark = True
changedMark = False
colorPallet = getColorPallet(10)
cv2.namedWindow("img", cv2.WINDOW_AUTOSIZE)
cv2.setMouseCallback("img", creatRoots)
colorMark = 0
while True:
key = cv2.waitKey(1) & 0xFF
if key == ord('q'):
break
elif chr(key).isdigit():
colorMark = eval(chr(key))
print(chr(key), colorMark)
if changedMark:
copied_markedImg = markedImg.copy()
cv2.watershed(srcImg, copied_markedImg)
segments = np.zeros_like(srcImg, np.uint8)
for clrIndex in range(len(colorPallet)):
segments[copied_markedImg == (clrIndex)] = colorPallet[clrIndex]
cv2.imshow("img", scene)
cv2.imshow("segments", segments)
cv2.destroyAllWindows()
def post_process(q=-5):
nc_logits = np.squeeze(
np.squeeze(
np.load(r'E:\MoNUSAC\MyModel\prediction\nc_logits.npy',
allow_pickle=True)))
nc_target = np.squeeze(
np.squeeze(
np.load(r'E:\MoNUSAC\MyModel\prediction\nc_target.npy',
allow_pickle=True)))
np_logits = np.squeeze(
np.squeeze(
np.load(r'E:\MoNUSAC\MyModel\prediction\np_logits.npy',
allow_pickle=True)))
hv_logits = np.squeeze(
np.load(r'E:\MoNUSAC\MyModel\prediction\hv_logits.npy',
allow_pickle=True))
image = np.squeeze(
np.load(r'E:\MoNUSAC\MyModel\prediction\image.npy', allow_pickle=True))
image = image.transpose((1, 2, 0))
image2 = image.copy()
image2 += 60 * cv2.merge([
np.where(nc_target == 1, 1, 0),
np.where(nc_target == 2, 2, 0) + np.where(nc_target == 4, 4, 0),
np.where(nc_target == 3, 3, 0) + np.where(nc_target == 4, 4, 0)
])
cv2.imshow('target', image2)
# cv2.imshow('h',np.squeeze(hv_logits[0,...]).astype(np.float))
h, v = cv2.Sobel(np.squeeze(hv_logits[0, :, :]),
ddepth=-1,
dx=1,
dy=0,
ksize=3), cv2.Sobel(np.squeeze(hv_logits[1, :, :]),
ddepth=-1,
dx=0,
dy=1,
ksize=3)
# show_('h',h)
# show_('v',v)
# cv2.imshow('H', h.astype(np.float))
# cv2.imshow('V', v.astype(np.float))
Sm = np.where(h < v, h, v)
t = np.argmax(np_logits, axis=0)
# t=np_logits
# cv2.imshow('sm', Sm.astype(np.float))
# show_('Sm',Sm)
tmp = t - np.where(Sm < q, 1, 0)
tmp = np.where(tmp > 0, 1, 0)
# cv2.imshow('tmp',tmp.astype(np.uint8)*255)
sure_bg = 1 - t
sure_fg = tmp
unkown = 1 - sure_bg - sure_fg
_, markers = cv2.connectedComponents(sure_fg.astype(np.uint8))
markers = markers + 1
markers[sure_bg == 1] = 0
# cv2.imshow('mark',markers.astype(np.uint8))
img = cv2.distanceTransform(t.astype(np.uint8), cv2.DIST_L2,
cv2.DIST_MASK_3)
img = (np.max(img) - img) / (np.max(img) - np.min(img)) * 255
img = img.astype(np.uint8)
img = cv2.merge([img, img, img])
# cv2.imshow('img',img)
markers = cv2.watershed(img, markers)
m = np.zeros((256, 256))
nc_logits = np.argmax(nc_logits, axis=0)
for i in np.unique(markers).tolist():
if i > 2:
temp = markers.copy()
temp = np.where(temp == i, 1, 0)
temp = temp * nc_logits
maxs = 0
maxc = 0
for j in np.unique(temp).tolist():
if j > 0:
temp2 = np.where(temp == j, 1, 0)
s = np.sum(temp2)
if s > maxs:
maxs = s
maxc = j
m += np.where(markers == i, maxc, 0)
image += 60 * cv2.merge([
np.where(m == 1, 1, 0),
np.where(m == 2, 2, 0) + np.where(m == 4, 4, 0),
np.where(m == 3, 3, 0) + np.where(m == 4, 4, 0)
])
image[markers == -1] = [0, 0, 255]
cv2.imshow('img', image)
cv2.waitKey(0)
def boundary_refinement_watershed(X, Y_pred, erode=True, save_marks_dir=None):
"""Apply watershed to the given predictions with the goal of refine the
boundaries of the artifacts.
Based on https://docs.opencv.org/master/d3/db4/tutorial_py_watershed.html.
Parameters
----------
X : 4D Numpy array
Original data to guide the watershed. E.g. ``(img_number, x, y, channels)``.
Y_pred : 4D Numpy array
Predicted data to refine the boundaries. E.g. ``(img_number, x, y, channels)``.
erode : bool, optional
To extract the sure foreground eroding the artifacts instead of doing
with distanceTransform.
save_marks_dir : str, optional
Directory to save the markers used to make the watershed. Useful for
debugging.
Returns
-------
Array : 4D Numpy array
Refined boundaries of the predictions. E.g. ``(img_number, x, y, channels)``.
Examples
--------
+-----------------------------------------+-----------------------------------------+
| .. figure:: img/FIBSEM_test_0.png | .. figure:: img/FIBSEM_test_0_gt.png |
| :width: 80% | :width: 80% |
| :align: center | :align: center |
| | |
| Original image | Ground truth |
+-----------------------------------------+-----------------------------------------+
| .. figure:: img/FIBSEM_test_0_pred.png | .. figure:: img/FIBSEM_test_0_wa.png |
| :width: 80% | :width: 80% |
| :align: center | :align: center |
| | |
| Predicted image | Watershed ouput |
+-----------------------------------------+-----------------------------------------+
The marks used to guide the watershed is this example are these:
.. image:: img/watershed2_marks_test0.png
:width: 70%
:align: center
"""
if save_marks_dir is not None:
os.makedirs(save_marks_dir, exist_ok=True)
watershed_predictions = np.zeros(Y_pred.shape[:3])
kernel = np.ones((3, 3), np.uint8)
d = len(str(X.shape[0]))
for i in tqdm(range(X.shape[0])):
im = cv2.cvtColor(X[i, ...] * 255, cv2.COLOR_GRAY2RGB)
pred = Y_pred[i, ..., 0]
# sure background area
sure_bg = cv2.dilate(pred, kernel, iterations=3)
sure_bg = np.uint8(sure_bg)
# Finding sure foreground area
if erode:
sure_fg = cv2.erode(pred, kernel, iterations=3)
else:
dist_transform = cv2.distanceTransform(a, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform,
0.7 * dist_transform.max(), 255, 0)
sure_fg = np.uint8(sure_fg)
# Finding unknown region
unknown_reg = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown_reg == 1] = 0
if save_marks_dir is not None:
f = os.path.join(save_marks_dir,
"mark_" + str(i).zfill(d) + ".png")
cv2.imwrite(f, markers)
markers = cv2.watershed((im).astype(np.uint8), markers)
watershed_predictions[i] = markers
# Label all artifacts into 1 and the background with 0
watershed_predictions[watershed_predictions == 1] = 0
watershed_predictions[watershed_predictions > 1] = 1
watershed_predictions[watershed_predictions == -1] = 0
return np.expand_dims(watershed_predictions, -1)
cv2.namedWindow('Road Image')
cv2.setMouseCallback('Road Image', mouse_callback)
while True:
cv2.imshow('Watershed Segments', segments)
cv2.imshow('Road Image', road_copy)
k = cv2.waitKey(1)
if k == 27:
break
elif k == ord('c'):
road_copy = road.copy()
marker_image = np.zeros(road.shape[:2], dtype=np.int32)
segments = np.zeros(road.shape, dtype=np.uint8)
elif k > 0 and chr(k).isdigit():
current_marker = int(chr(k))
if marks_updated:
marker_image_copy = marker_image.copy()
cv2.watershed(road, marker_image_copy)
segments = np.zeros(road.shape, dtype=np.uint8)
for color_ind in range(n_markers):
segments[marker_image_copy == (color_ind)] = colors[color_ind]
marks_updated = False
cv2.destroyAllWindows()
sx = np.min(ind[0]) - border_width
ex = np.max(ind[0]) + border_width
sy = np.min(ind[1]) - border_width
ey = np.max(ind[1]) + border_width
# distance transform
one_comp = oldmarkers[sx:ex, sy:ey] == i
compon_dis = cv2.distanceTransform(one_comp.astype(np.uint8), cv2.DIST_L2,
5)
_, dis_peak = cv2.threshold(compon_dis, 0.75 * compon_dis.max(), 255, 0)
# watershed
ret, new_markers = cv2.connectedComponents(dis_peak.astype(np.uint8))
new_markers = new_markers + 1
new_markers[(one_comp == True) & (new_markers == 1)] = 0
water = cv2.watershed(img[sx:ex, sy:ey], new_markers.copy())
# draw
for j in range(2, ret + 1):
c = cv2.findContours(np.uint8(water == j), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)[1]
if len(c):
c = c[0]
else:
continue
c[:, 0, 0] += sy
c[:, 0, 1] += sx
# cv2.drawContours(img, [c], -1, (0, 0, 255), 5)
mm = 20
minx = np.min(c[:, 0, 0]) - mm
maxx = np.max(c[:, 0, 0]) + mm
D2 = ndi.distance_transform_edt(fat_markers)
dist_binary2 = D2 > dt_t
dist_binary2 = dist_binary2.astype(np.float)
# plt.figure();plt.imshow(dist_binary2)
dist_binary2_ = morphology.remove_small_holes(
util.pad(dist_binary2, (1, ), 'constant',
constant_values=0).astype(np.int), 300)
dist_binary2_ = dist_binary2_[1:-1, 1:-1]
# plt.figure();plt.imshow(dist_binary2_)
sure_fg = dist_binary1_.astype(np.float) * dist_binary2_.astype(np.float)
sure_fgo = morphology.binary_opening(sure_fg, morphology.disk(5))
unknown = np.maximum(sure_bg - sure_fgo, 0)
markers = measure.label(sure_fgo, background=0)
markers += 1
markers[unknown == 1] = 0
# plt.figure();plt.imshow(markers, cmap=cmap)
imgc_ = np.uint8(imgc * 255.)
labels = cv2.watershed(imgc_, markers)
plt.figure()
plt.imshow(labels.get(), cmap=cmap)
imgc_overlay = imgc
imgc_overlay[labels.get() == -1] = [1, 0, 0]
plt.figure()
plt.imshow(imgc_overlay)
#######################################################################
fg = compare(road_lbp, sample=(880, 100), high=64, width=32, condi=0.84)
# print(np.unique(fg, return_counts=True))
cv2.imshow("fg", fg)
# watershed
blur = cv2.GaussianBlur(gray, (5, 5), 0)
_, thres = cv2.threshold(blur, 0, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY_INV)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
mb = cv2.morphologyEx(thres, cv2.MORPH_OPEN, kernel, iterations=2)
bg = cv2.dilate(mb, kernel, iterations=3)
# cv2.imshow("bg", bg)
unknown = cv2.subtract(bg, fg)
_, markers = cv2.connectedComponents(fg)
# print(np.unique(markers, return_counts=True))
markers = cv2.watershed(road, markers=markers)
# markers = markers + 1
# markers[unknown == 255] = 4
road[markers == -1] = [0, 0, 0]
# print(road[markers == 1])
# cv2.imshow("watershed", road)
# draw
color_block = np.zeros(road.shape, dtype=np.uint8)
color_block[markers == 0] = [255, 255, 255]
color_block[markers == 1] = [191, 62, 255]
color_block[markers == 2] = [130, 130, 130]
color_block[markers == 3] = [152, 251, 152]
# cv2.imshow("color_block", color_block)
# result
def getWatershed(data, w):
height = np.array(data)
image = np.zeros(height.shape)
for i in range(len(data)):
for j in range(len(data[0])):
height[i][j] = data[i][j].elevation
maximum = np.amax(height)
minimum = np.amin(height)
numberOfBits = 16
interval = (maximum - minimum) / (pow(2, numberOfBits) - 1)
for i in range(len(height)):
for j in range(len(height[0])):
image[i][j] = int(data[i][j].getRisk(minimum, maximum) *
(pow(2, numberOfBits) - 1))
image = image.astype('uint8') * 255
cv2.imshow("w", resizeimage(image, 1000))
cv2.waitKey(0)
# image = cv2.bitwise_not(image)
# image = resizeimage(image, 1000)
image3 = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
ret, thresh = cv2.threshold(image, 0, 255,
cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
# noise removal
kernel = np.ones((3, 3), np.uint8)
# sure background area
sure_bg = cv2.dilate(thresh, kernel, iterations=3)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(thresh, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.3 * dist_transform.max(),
255, 0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
markers = cv2.watershed(image3, markers)
diff = []
for m in markers:
for a in m:
if not a in diff:
diff.append(a)
image3[markers == -1] = [255, 0, 0]
# Map to store region wise location objects
regions = getRegions(markers, data)
# Map to store region boundary location objects
regionBoundary = findboundary2(markers, data)
# Map to store regions and there relative sizes
sizeFactor = {}
# Size list
sizeList = {}
# Map to Store Average Depths
avgDepth = {}
# Map to store the centers of the regions
centers = {}
maxSize = 0
for k in regions:
# Find maximum size region
if k not in sizeFactor:
sizeFactor[k] = float(len(regions[k]))
maxSize = max(maxSize, sizeFactor[k])
sizeList[k] = len(regions[k])
# Find the center of the region i,e, the minimum depth region
minDepth = regions[k][0]
# Depth wise classification
depthSum = 0
for val in regions[k]:
if minDepth.elevation > val.elevation:
minDepth = val
depthSum = depthSum + val.elevation
centers[k] = minDepth
# Normalise the depth
avgDepth[k] = depthSum / len(regions[k])
# Normalize the size factors with respect to greatest size region
for k in sizeFactor:
sizeFactor[k] = sizeFactor[k] / maxSize
maxDepth = avgDepth[list(avgDepth.keys())[0]]
for k in avgDepth:
maxDepth = max(maxDepth, avgDepth[k])
for k in avgDepth:
avgDepth[k] = avgDepth[k] / maxDepth
factor1 = 0.5
factor2 = 0.5
riskFactor = {}
for k in regions:
riskFactor[k] = factor1 * (1 - avgDepth[k]) + factor2 * sizeFactor[k]
riskObjects = []
rs = []
for k in regions:
if k == -1:
continue
risk = riskFactor[k]
area = pow((pow(sizeList[k], 0.5) * w) / 1000, 2)
riskObjects.append(
RiskMapRegion(centers[k], regionBoundary[k], risk, area))
# rs.append((-risk, -area, k))
# rs.sort()
# rss = rs[:w]
# rb = {}
# for r in rss:
# rb[r[2]] = regionBoundary[r[2]]
# print("YO")
# sb = sort_boundaries(rb, data)
# for r in rss:
# riskObjects.append(RiskMapRegion(centers[r[2]],sb[r[2]],risk,area))
return sorted(riskObjects, key=func)
def find_contours(image):
# Conversion to HSV space and taking the saturation channel
hsv = cv.cvtColor(image, cv.COLOR_BGR2HSV)
img = cv.split(hsv)[1]
# cv.imwrite("0_saturation.png", img)
# Thresholding
_, dst = cv.threshold(img, 0, 255, cv.THRESH_BINARY + cv.THRESH_OTSU)
# cv.imwrite("1_thresholded.png", dst)
# Mask borders
b_mask = borders_mask(image)
# cv.imwrite("2_borders_mask.png", b_mask)
dst = cv.bitwise_and(dst, b_mask)
# Finding contours to remove holes
contours, _ = cv.findContours(dst, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_NONE)
# tmp = np.zeros((400, 600), dtype=np.uint8)
# cv.drawContours(tmp, contours, -1, (255, 255, 255), -1, 4)
# cv.imwrite("3_unfiltered_cc.png", tmp)
final = None
final_ref = 1.0
for c in contours:
area = cv.contourArea(c)
perimeter = cv.arcLength(c, True)
circ = circularity(area, perimeter)
if area > 400 and circ < 0.8:
if area > final_ref:
final = c
final_ref = area
mask = np.zeros((400, 600), dtype=np.uint8)
if final is not None:
cv.drawContours(mask, [final], -1, (255, 255, 255), -1, 4)
# cv.imwrite("4_final_mask.png", mask)
### WATERSHED START
## FOREGROUND
# Distance transform and thresholding
fg = cv.distanceTransform(mask, cv.DIST_L2, cv.DIST_MASK_PRECISE)
fg_min = np.amin(fg)
fg_max = np.amax(fg)
fg[:] = (fg - fg_min) * 255 // (fg_max - fg_min)
fg = np.uint8(fg)
_, fg = cv.threshold(fg, 100, 255, cv.THRESH_BINARY)
## BACKGROUND
bg = cv.dilate(mask, None, iterations=4)
bg[bg == 0] = 128
bg[bg == 255] = 0
## MARKERS
markers = cv.add(fg, bg)
# cv.imwrite("5_watershed_mask.png", markers)
markers32 = np.int32(markers)
cv.watershed(image, markers32)
w_mask = cv.convertScaleAbs(markers32)
# cv.imwrite("6_watershed_mask.png", w_mask)
w_mask[w_mask == 128] = 0
w_mask[w_mask == 1] = 255
w_mask[0, :] = 0
w_mask[w_mask.shape[0] - 1, :] = 0
w_mask[:, 0] = 0
w_mask[:, w_mask.shape[1] - 1] = 0
# cv.imwrite("7_final_result.png", w_mask)
return w_mask
def bottleCapDetect(img):
scale_factor = min(1024 / img.shape[0], 768 / img.shape[1], 1)
img = cv2.resize(img, (0, 0), fx=scale_factor, fy=scale_factor)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
edge_threshold = 10
edeg_threahold_max = 2 * edge_threshold
blurred = cv2.medianBlur(gray, 3)
edges = cv2.Canny(blurred,
edge_threshold,
edeg_threahold_max,
apertureSize=3)
struct_ele = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
edges_closed = edges.copy()
edges_closed = cv2.morphologyEx(edges_closed,
cv2.MORPH_DILATE,
struct_ele,
iterations=1)
edges_closed = cv2.morphologyEx(edges_closed,
cv2.MORPH_CLOSE,
struct_ele,
iterations=5)
mask = np.zeros((edges_closed.shape[0] + 2, edges_closed.shape[1] + 2),
dtype="uint8")
cv2.floodFill(edges_closed,
mask, (0, 0),
255,
flags=cv2.FLOODFILL_MASK_ONLY)[1]
mask[...] = 255 * (1 - mask)
cap_cnt = 10
def get_markers(mask_orig):
mask = mask_orig.copy()
cnt = 0
result = []
while cnt < cap_cnt:
mask = cv2.medianBlur(mask, 9)
# do distance transform
dist = cv2.distanceTransform(mask, cv2.DIST_L2, 5)
dist = dist.astype("uint8")
ret, markers_binary = cv2.threshold(dist, 0.8 * dist.max(), 255, 0)
# do marker labelling
ret, markers = cv2.connectedComponents(markers_binary)
cur_cnt = markers.max()
print("Got", cur_cnt, "marker(s)")
cnt += cur_cnt
cur_result = []
for i in range(1, cur_cnt + 1):
pos = np.nonzero(markers == i)
x, y = pos[1], pos[0]
minx, maxx, miny, maxy = x.min(), x.max(), y.min(), y.max()
w = np.max(dist[markers == i])
cur_result.append(((minx, miny), (maxx, maxy), w))
result.extend(cur_result)
if cnt < cap_cnt:
for i in range(cur_cnt):
(minx, miny), (maxx, maxy), w = cur_result[i]
radius = w + 20 # just in case :)
print("Removing", (minx, miny), (maxx, maxy), radius)
mask = cv2.circle(mask,
((minx + maxx) // 2, (miny + maxy) // 2),
radius, 0, -1)
elif cnt > cap_cnt:
print("warning: 翻车啦")
return result
markers = get_markers(mask)
# now prepare for watersheding
ws = np.logical_not(mask).astype('int32')
for i, (p1, p2, w) in enumerate(markers):
center = (p1[0] + p2[0]) // 2, (p1[1] + p2[1]) // 2
axis = ((p2[0] - p1[0]) // 2, (p2[1] - p1[1]) // 2)
cv2.ellipse(ws, center, axis, 0, 0, 360, i + 2, cv2.FILLED)
flooded = ws.copy()
flooded = cv2.watershed(cv2.cvtColor(mask, cv2.COLOR_GRAY2BGR), flooded)
flooded[...] = flooded - 1
flooded[flooded <= 0] = 0
# dilate a little to remove tiny edges
flooded = flooded.astype("uint8")
flooded = cv2.morphologyEx(flooded,
cv2.MORPH_DILATE,
struct_ele,
iterations=1)
ws = flooded
boxed = img.copy()
bounding_boxes = []
minimal_bounding_boxes = []
for i in range(cap_cnt):
# value in marker image is i+1
buf = (ws == i + 1).astype('uint8')
contours, hierarchy = cv2.findContours(buf, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)
assert (len(contours) == 1)
x, y, w, h = cv2.boundingRect(contours[0])
bounding_boxes.append(((x, y), (x + w, y + h)))
minimal_bounding_boxes.append(cv2.minAreaRect(contours[0]))
ellipse = cv2.fitEllipse(contours[0])
cv2.ellipse(boxed, ellipse, (0, 255, 0), 2)
box_points = cv2.boxPoints(minimal_bounding_boxes[-1])
box_points = np.int0(box_points)
cv2.rectangle(boxed, bounding_boxes[-1][0], bounding_boxes[-1][1],
(255, 0, 0), 2)
cv2.drawContours(boxed, [box_points], 0, (0, 0, 255), 2)
real_caps = []
cap_edges = []
real_ans = []
cord = []
edges = cv2.Canny(img, 40, 80)
for i, ((p1, p2),
(center, (width, height),
a)) in enumerate(zip(bounding_boxes, minimal_bounding_boxes)):
g = gray[p1[1]:p2[1], p1[0]:p2[0]]
e = edges[p1[1]:p2[1], p1[0]:p2[0]]
cord.append(center)
if width / height > 1.5 or height / width > 1.5:
print(i, "is a side")
real_ans.append('side')
else:
real_ans.append(None)
# do circle approximation for better center mounting
circles = cv2.HoughCircles(g, cv2.HOUGH_GRADIENT, 2, 40, 1, 20, 40,
200)
if type(circles[0][0]) == np.ndarray:
for i in circles[0, :]:
# draw the outer circle
cv2.circle(g, (i[0], i[1]), i[2], 255, 2)
# draw the center of the circle
cv2.circle(g, (i[0], i[1]), 2, 255, 3)
real_caps.append(g)
cap_edges.append(e)
k = len(real_caps)
t = cv2.getGaussianKernel(9, 1)
gaussian_kernel = t * t.T
gaussian_kernel *= 255 / np.max(gaussian_kernel)
gaussian_kernel = np.uint8(gaussian_kernel)
gaussian_kernel = cv2.resize(gaussian_kernel, (64, 64))
for no, (cap, edge) in enumerate(zip(real_caps, cap_edges)):
if real_ans[no] is not None:
print("Skipping", no, "it is a side")
continue
c = cv2.resize(cap, (128, 128))
e = cv2.resize(edge, (128, 128))
rx, ry = c.shape[0] // 4, c.shape[1] // 4
cx, cy = c.shape[0] // 2, c.shape[1] // 2
score = np.sum(e[cx - rx:cx + rx, cy - ry:cy + ry] * gaussian_kernel)
print(score)
real_ans[no] = 'top' if score > 50000 else 'tail'
labelled = img.copy()
for i in range(len(real_ans)):
if real_ans[i] == 'top':
color = (255, 0, 0)
elif real_ans[i] == 'tail':
color = (0, 255, 0)
else:
color = (0, 0, 255)
cv2.rectangle(labelled, bounding_boxes[i][0], bounding_boxes[i][1],
color, 2)
cv2.putText(labelled, real_ans[i], bounding_boxes[i][0],
cv2.FONT_HERSHEY_SIMPLEX, 1, color, 2)
cv2.putText(labelled, "({},{})".format(int(cord[i][0]),
int(cord[i][1])),
(int(cord[i][0]) - 50, int(cord[i][1])),
cv2.FONT_HERSHEY_SIMPLEX, 0.7, color, 2)
return labelled
def segment_by_watershed(
binary_img: np.ndarray,
img: np.ndarray,
dist_thresh: float = 3.0) -> Tuple[np.ndarray, List[np.ndarray]]:
"""Segment image using watershed algorithm.
Parameters
----------
binary_img:np.ndarray
Binary image derived from ``img`` with nonzero pixel blobs for objects.
This is usually produced after converting the ``img`` to grayscale and
then thresholding.
img: np.ndarray
Original image to be segmented.
dist_thresh: float, optional
Threshold for distance of pixels from boundary to consider them core
points. If it is < 1.0, it is interpreted as fraction of the maximum
of the distances.
Returns
-------
points_list: list[np.ndarray]
List of arrays containing positions of the pixels in each object.
Notes
-----
This code is derivative of this OpenCV tutorial:
https://docs.opencv.org/master/d3/db4/tutorial_py_watershed.html
and falls under the same licensing.
"""
kernel = np.ones((3, 3), dtype=np.uint8)
opening = cv2.morphologyEx(binary_img,
cv2.MORPH_OPEN,
kernel,
iterations=2)
# Distance transform calculates the distance of each pixel from
# background (black in this case) pixels. So we have an image
# where pixel intensity means the distance of that pixel from the
# background
dist_xform = cv2.distanceTransform(opening, cv2.DIST_L2,
cv2.DIST_MASK_PRECISE)
# Thresholding the distance image to find pixels which are more
# than a certain distance away from background - this should give
# us the pixels central to foreground objects
if dist_thresh < 1.0:
# threshold relative to maximum of computed distance
dist_thresh *= dist_xform.max()
ret, sure_fg = cv2.threshold(dist_xform, dist_thresh, 255, 0)
sure_fg = np.uint8(sure_fg)
# border between background and foreground
sure_bg = cv2.dilate(opening, kernel, iterations=3)
unknown = cv2.subtract(sure_bg, sure_fg)
ret, markers = cv2.connectedComponents(sure_fg, connectivity=4)
# logging.debug(f'Found {ret} connected components')
# 0 is for background - assign a large value to keep them off
markers += 1
markers[unknown == 255] = 0
markers = cv2.watershed(img, markers)
unique_labels = set(markers.flat)
unique_labels.discard(-1)
unique_labels.discard(0)
ret = [np.argwhere(markers == label) for label in sorted(unique_labels)]
# markers[markers == -1] = 0
# markers = np.uint8(markers)
# Fast swapping of y and x - see answer by blax here:
# https://stackoverflow.com/questions/4857927/swapping-columns-in-a-numpy-array
for points in ret:
points[:, 0], points[:, 1] = points[:, 1], points[:, 0].copy()
return ret
def get_track(img):
"""
Creates and returns a binary image of the track. Also finds track sections and starts the calculation of the racing line
:param img: the image
:return: a binary image of the track
"""
# Set the bounds to detect the black parts of the track and invert it to get the actual track
lower_bounds_track = np.array([0, 0, 0])
upper_bounds_track = np.array([10, 10, 10])
track = cv.bitwise_not(
cv.inRange(img, lower_bounds_track, upper_bounds_track))
# Closing (Dilate + Erode) to remove noise (of the ball border)
track = cv.morphologyEx(track, cv.MORPH_CLOSE, np.ones((5, 5), np.uint8))
# Find the finish line
lower_bounds_finish = np.array([200, 100, 0])
upper_bounds_finish = np.array([255, 150, 10])
finish_line = cv.dilate(cv.inRange(img, lower_bounds_finish,
upper_bounds_finish),
np.ones((5, 5), np.uint8),
iterations=1)
finish_line_coords = cv.findNonZero(finish_line)
finish_line_center = np.mean(finish_line_coords, axis=0)[0]
# Get the contours of the track
_, contours, _ = cv.findContours(track, cv.RETR_TREE, cv.CHAIN_APPROX_NONE)
image = track.copy()
# Get the inner and outer contour and the indices of the points where they touch the finish line
inner_contour, outer_contour, inner_start_index, outer_start_index, inner_contour_index = _get_inner_and_outer_contour(
contours, finish_line_coords)
# Calculate the starting direction based on the center of the actual start line and center of the orange area
direction = finish_line_center - MatrixOps.convex_combination(
inner_contour[inner_start_index], outer_contour[outer_start_index],
0.5)
direction = direction / MatrixOps.fast_norm(direction)
### Calculate cross sections of the track ###
# Create empty/black 3-channel image to run watershed on
img_empty = np.zeros(img.shape, np.uint8)
# Create a 1-channel image for the markers
markers = np.zeros(image.shape, np.int32)
# Mark the inner contour with label 1 and the other contours with label 2 by drawing them on the markers image
for i in range(0, len(contours)):
cv.drawContours(markers, contours, i,
1 if i == inner_contour_index else 2)
# Use watershed to get the boundary between inner and outer contour
markers = cv.watershed(img_empty, markers)
# Remove '-1' markers on the border of the image
markers[:, 0] = markers[1, 1]
markers[0, :] = markers[1, 1]
markers[:, -1] = markers[1, 1]
markers[-1, :] = markers[1, 1]
# Get a sorted vector of all positions of points on the watershed border
_, watershed_contours, _ = cv.findContours(
(markers == -1).astype(np.uint8), cv.RETR_TREE, cv.CHAIN_APPROX_NONE)
watershed_contour = watershed_contours[0].reshape(
len(watershed_contours[0]), 2)
img[markers == -1] = [0, 0, 255]
# cv.imshow("color", img) TODO remove #
# Circshift the contour so that the first point corresponds to the start/finishing line
start_index = MatrixOps.find_closest_point_index(
MatrixOps.convex_combination(inner_contour[inner_start_index],
outer_contour[outer_start_index], 0.5),
watershed_contour)
watershed_contour = np.roll(watershed_contour, -2 * start_index)
# Check if flipping the contour array is necessary based on the start direction
next_point = watershed_contour[0] + direction
closest_point_index = MatrixOps.find_closest_point_index(
next_point, watershed_contour[[1, len(watershed_contour) - 1]])
if closest_point_index == 1:
watershed_contour = np.flip(watershed_contour, 0)
# Reduce the number of points on the contour to get the checkpoints needed for smoothing and for the features
checkpoints = watershed_contour[::20]
#for c in checkpoints:
# cv.circle(image, tuple(c), 3, (192, 192, 192), -1)
#cv.imshow("image", image)
# Erode the track/dilate the track border by the radius of the ball
kernel = np.ones((5, 5), np.uint8)
eroded_track = cv.erode(
image, kernel, iterations=4
) # Kernel increases border by 2 pixel each iteration, ball seems to have a radius of about 8 pixels
#cv.imshow("eroded_track", eroded_track) TODO remove #
smoothened_checkpoints = _smoothen_checkpoints_contour(
checkpoints, eroded_track)
#for c in smoothened_checkpoints:
# cv.circle(image, tuple(c), 3, (192, 192, 192), -1)
#cv.imshow("image", image)
return track, smoothened_checkpoints
DistNorm = np.uint8(DistImg * C)
cv2.imshow('DistNorm', DistNorm)
_, DistBin = cv2.threshold(src=DistNorm,
thresh=DistNorm.max() * 0.7,
maxval=255,
type=cv2.THRESH_BINARY)
cv2.imshow('DistBin', DistBin)
# cv2.imshow('DistImg', cv2.normalize(src=DistImg, norm_type=cv2.NORM_MINMAX))
# unknown
# Unknown = DilationImg - DistBin
Unknown = cv2.subtract(DilationImg, DistBin)
cv2.imshow('Unknown', Unknown)
# connect
MarkImg = np.zeros(GrayImg.shape, dtype=np.int32)
MarkImg[DilationImg == 0] = 1
ForegroundContours, _ = cv2.findContours(image=DistBin.copy(),
mode=cv2.RETR_EXTERNAL,
method=cv2.CHAIN_APPROX_NONE)
ret, MarkImg = markContours(ForegroundContours,
markImg=MarkImg,
startNum=2)
cv2.imshow('MarkImg', np.uint8(MarkImg * (255.0 / MarkImg.max())))
cv2.watershed(SrcImg, MarkImg)
SrcImg[MarkImg == -1] = [0, 0, 255]
cv2.imshow('Src', SrcImg)
cv2.imshow('Gray', GrayImg)
cv2.imshow('Bin', BinImg)
cv2.waitKey()
def textDetectWatershed(thresh, original, draw=False, group=False):
""" Text detection using watershed algorithm """
# According to: http://docs.opencv.org/trunk/d3/db4/tutorial_py_watershed.html
img = resize(original, 3000)
thresh = resize(thresh, 3000)
# noise removal
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=3)
# sure background area
sure_bg = cv2.dilate(opening, kernel, iterations=3)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.001 * dist_transform.max(),
255, 0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers += 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
markers = cv2.watershed(img, markers)
image = img.copy()
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# Creating result array
boxes = []
for mark in np.unique(markers):
# mark == 0 --> background
if mark == 0:
continue
# Draw it on mask and detect biggest contour
mask = np.zeros(gray.shape, dtype="uint8")
mask[markers == mark] = 255
cnts = cv2.findContours(mask.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)[-2]
c = max(cnts, key=cv2.contourArea)
# Draw a bounding rectangle if it contains text
x, y, w, h = cv2.boundingRect(c)
cv2.drawContours(mask, c, 0, (255, 255, 255), cv2.FILLED)
maskROI = mask[y:y + h, x:x + w]
# Ratio of white pixels to area of bounding rectangle
r = cv2.countNonZero(maskROI) / (w * h)
# Limits for text
if r > 0.1 and 2000 > w > 15 and 1500 > h > 15:
boxes += [[x, y, w, h]]
# Group intersecting rectangles
if group:
boxes = group_rectangles(boxes)
boxes.sort(key=lambda b: (b[1] + b[3], b[0]))
bounding_boxes = np.array([0, 0, 0, 0])
rois = []
for (x, y, w, h) in boxes:
rois.append(img[y:y + h, x:x + w])
cv2.rectangle(image, (x, y), (x + w, y + h), (0, 255, 0), 2)
bounding_boxes = np.vstack(
(bounding_boxes, np.array([x, y, x + w, y + h])))
if draw:
implt(sure_bg, t='sure_background')
implt(sure_fg, t='sure_foreground')
implt(unknown, t='unknown area')
implt(markers, t='Markers')
implt(image)
# Recalculate coordinates to original size
boxes = bounding_boxes[1:].dot(ratio(original,
img.shape[0])).astype(np.int64)
return boxes, rois
def process(filename, output_folder, verbose):
img = cv.imread(filename)
grayscale = cv.imread(filename, 0)
filename = os.path.basename(filename).split('.')[-2]
# Edge enchancement
unsharp = unsharp_mask(grayscale, amount=6, threshold=0)
bilateral_blur = cv.bilateralFilter(unsharp, 9, 75, 75)
_, th = cv.threshold(bilateral_blur, 0, 255, cv.THRESH_BINARY+cv.THRESH_OTSU)
# Noise reduction
kernel = np.ones((3,3), np.uint8)
m_blur = cv.medianBlur(cv.medianBlur(th, 3), 3)
opening = cv.morphologyEx(m_blur, cv.MORPH_OPEN, kernel, iterations = 1)
# Watershed segmentation
sure_bg = cv.dilate(opening, kernel, iterations=2)
dist_transform = cv.distanceTransform(opening, cv.DIST_L2, 0)
dist_transform = cv.normalize(dist_transform, dist_transform, 0, 255, cv.NORM_MINMAX)
_, sure_fg = cv.threshold(dist_transform, 0.23 * dist_transform.max(), 255, 0)
sure_fg = np.uint8(sure_fg * 100000)
unknown = cv.subtract(sure_bg, sure_fg)
_, markers = cv.connectedComponents(sure_fg)
markers = markers + 1
markers[unknown==255] = 0
markers = cv.watershed(img, markers)
bordered = opening
bordered[markers == -1] = 0
bordered = cv.convertScaleAbs(bordered)
bordered = cv.erode(bordered, kernel, iterations=1)
# Contour detection
contours, hierarchy = cv.findContours(bordered, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
contours = [c for c in contours if cv.contourArea(c)>5] # Filtering erroneous contours (with tiny areas)
moments = [cv.moments(c) for c in contours]
centroids = [(int(m['m10']/m['m00']), int(m['m01']/m['m00'])) for m in moments]
# Scaling contours up (to revert errosion effect)
contours = [np.asarray(((c - cent)*1.05)+cent, int) for c, cent in zip(contours, centroids)]
result = cv.cvtColor(grayscale, cv.COLOR_GRAY2RGB)
cv.drawContours(result, contours, -1, (0,255,0), thickness=1, lineType=cv.LINE_AA)
contours_filled = np.zeros(grayscale.shape)
cv.fillPoly(contours_filled, pts=contours, color=(255,255,255))
# Calculation of equivalent diameter
equivalent_diameters = [np.sqrt(4 * m['m00'] / np.pi) for m in moments]
max_ = max(equivalent_diameters)
min_ = min(equivalent_diameters)
contours_colored = cv.cvtColor(grayscale, cv.COLOR_GRAY2RGB)
for c, d in zip(contours, equivalent_diameters):
color = (int(255*(d-min_)/(max_-min_)), 0, int(255-255*(d-min_)/(max_-min_)))
cv.drawContours(contours_colored, [c], -1, color, thickness=2, lineType=cv.LINE_AA)
colormap = colors.LinearSegmentedColormap.from_list("BuRd", [(0,0,1), (1,0,0)], 256)
plt.figure(figsize=(10,10))
plt.axis('off')
plt.title('Particles colored with respect to equivalent diameter.')
plt.imshow(contours_colored, colormap)
cbar = plt.colorbar(fraction=0.0455, pad=0.04)
interp = lambda x: (int(x)*(max_-min_)+255*min_)//255
cbar.ax.set_yticklabels([interp(l.get_text()) for l in cbar.ax.get_yticklabels()])
plt.savefig(output_folder+filename+'_contours_colored.png', dpi=200)
# Size distribution
plt.figure(figsize=(10,3))
plt.title('Particle size distribution.')
nums, bins, patches = plt.hist(equivalent_diameters, bins=16)
bins /= max(bins)
for b, p in zip(bins, patches):
plt.setp(p, 'facecolor', colormap(b))
plt.savefig(output_folder+filename+'_size_distribution.png', dpi=150)
cv.imwrite(output_folder+filename+'_contours_filled.png', contours_filled)
if verbose:
cv.imwrite(output_folder+filename+'_unsharp.png', unsharp)
cv.imwrite(output_folder+filename+'_th.png', th)
cv.imwrite(output_folder+filename+'_opening.png', opening)
cv.imwrite(output_folder+filename+'_sure_bg.png', sure_bg)
cv.imwrite(output_folder+filename+'_dist_transform.png', dist_transform)
cv.imwrite(output_folder+filename+'_sure_fg.png', sure_fg)
cv.imwrite(output_folder+filename+'_unknown.png', unknown)
cv.imwrite(output_folder+filename+'_markers.png', markers)
cv.imwrite(output_folder+filename+'_contours.png', result)
foreground = foreground.astype(np.uint8)
#label是标签，该函数把图像背景标成0，其他封闭目标用从1开始的整数标记
#stats 是bounding box的信息，N*5的矩阵，行对应每个label，五列分别为[x0, y0, 宽度, 高度, 面积]
#centroids 是每个域的质心坐标
_, labels, stats, centroids = cv2.connectedComponentsWithStats(foreground)
#使背景标记为1
labels = labels + 1
ax4 = fig.add_subplot(164)
ax4.imshow(labels)
ax4.set_title("foreground")
#确定的前景目标标记为2、3、4......(不同目标标记为不同序号，方面后面进行粘连前景的分割)
markers[np.where(labels != 1)] = labels[np.where(labels != 1)]
markers2 = markers.copy()
ax5 = fig.add_subplot(165)
ax5.imshow(markers)
ax5.set_title("markers")
#只接受三通道的图像
#分水岭变换的结果会保存在markers2中
cv2.watershed(img, markers2)
ax6 = fig.add_subplot(166)
ax6.imshow(markers2)
ax6.set_title("obj")
plt.show()
def contours_watershed_filter(frame, contours):
list_final_Contourn = []
frame_height, frame_width = frame.shape[:2]
i = 0
for cnt in contours:
black_frame = np.zeros((frame_height, frame_width, 1), dtype="uint8")
cv2.drawContours(black_frame, [cnt], 0, 255, -1)
x_rect, y_rect, w, h = cv2.boundingRect(cnt)
x_prec = int(round(x_rect - (w / 2)))
x2_prec = x_prec + (2 * w)
y_prec = int(round(y_rect - (h / 2)))
y2_prec = y_prec + (2 * h)
height, width = frame.shape[:2]
x = 0 if x_prec < 0 else x_prec
y = 0 if y_prec < 0 else y_prec
x2 = width - 1 if x2_prec > width - 1 else x2_prec
y2 = height - 1 if y2_prec > height - 1 else y2_prec
imCrop = frame[y:y2, x:x2]
cropCopy = imCrop.copy()
mask_crop = black_frame[y:y2, x:x2]
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(mask_crop,
cv2.MORPH_OPEN,
kernel,
iterations=1)
sure_bg = cv2.dilate(opening, kernel, iterations=5)
sure_fg = np.uint8(opening)
unknown = cv2.subtract(sure_bg, sure_fg)
ret, markers = cv2.connectedComponents(sure_fg)
markers = markers + 1
markers[unknown == 255] = 0
markers = cv2.watershed(cropCopy, markers)
markers[:, 0] = 1
markers[0, :] = 1
markers[-1, :] = 1
markers[:, -1] = 1
cropCopy[markers == -1] = [255, 0, 0]
black_little_frame = np.zeros((y2 - y, x2 - x), np.uint8)
black_little_frame[markers == -1] = 255
contours_little, hieracy = cv2.findContours(black_little_frame,
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)
img_contours_little = np.zeros(
(black_little_frame.shape[0], black_little_frame.shape[1], 1),
dtype="uint8")
cv2.drawContours(img_contours_little, contours_little, -1, 255, -1)
momenti = cv2.moments(img_contours_little, True)
c = ((momenti['m00'])**2) / (2 * math.pi *
(momenti['mu20'] + momenti['mu02']))
if c >= 0.95:
hu_moments = cv2.HuMoments(momenti).flatten()
list_final_Contourn.append(cnt)
i += 1
return list_final_Contourn
def crop_img(img, indx, done=False):
if done is False:
plt.imshow(img)
plt.show()
#sets the background of the image to black
mask = np.zeros(img.shape[:2], np.uint8)
bgdModel = np.zeros((1, 65), np.float64)
fgdModel = np.zeros((1, 65), np.float64)
rect = (50, 50, 450, 290)
cv2.grabCut(img, mask, rect, bgdModel, fgdModel, 5,
cv2.GC_INIT_WITH_RECT)
mask2 = np.where((mask == 2) | (mask == 0), 0, 1).astype('uint8')
img = img * mask2[:, :, np.newaxis]
plt.imshow(img)
#plt.show()
#picks out a specific object in the image and crops out everything but that
#code from opencv's website on watershed
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
ret, thresh = cv2.threshold(gray, 100, 255,
cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
# noise removal
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=2)
# sure background area
sure_bg = cv2.dilate(opening, kernel, iterations=3)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.7 * dist_transform.max(),
255, 0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
markers = cv2.watershed(img, markers)
img[markers == -1] = [255, 255, 255]
loc = []
plt.imshow(img)
plt.show()
#performs the cropping of the image
for i in range(len(markers)):
for j in range(len(markers[i])):
if markers[i][j] == 1:
loc.append((i, j))
if len(loc) > 1:
loc.pop(0)
x = []
for i in range(len(loc)):
x.append(loc[i][0])
y = []
for i in range(len(loc)):
y.append(loc[i][1])
minx = np.min(x) - 3
miny = np.min(y) - 3
maxx = np.max(x) + 3
maxy = np.max(y) + 3
img[markers == -1] = [255, 0, 0]
plt.imshow(img)
plt.show()
roi = img[minx:maxx, miny:maxy]
#roi_rgb = cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)
plt.imshow(roi)
plt.show()
cv2.imwrite("roi.jpg", roi)
while True:
cv2.imshow('Image',selectOnImage_copy)
cv2.imshow('Watershed Segments', segments)
k = cv2.waitKey(1)
if k == 27:
break
elif k == ord('c'):
selectOnImage_copy = image.copy()
marker_image = np.zeros(image.shape[:2],dtype=np.int32)
segments = np.zeros(image.shape, dtype=np.uint8)
elif k>0 and chr(k).isdigit():
current_marker = int(chr(k))
if marks_updated:
marker_image_copy = marker_image.copy()
cv2.watershed(image,marker_image_copy)
segments = np.zeros(image.shape,dtype=np.uint8)
for color_ind in range(n_markers):
segments[marker_image_copy==(color_ind)] = colors[color_ind]
cv2.imwrite('images/seg.jpg',segments)
cv2.destroyAllWindows()
cimg = cv2.imread("images/seg.jpg", cv2.IMREAD_GRAYSCALE)
circles = cv2.HoughCircles(cimg,cv2.HOUGH_GRADIENT,1.5,120,param1=50,param2=30,minRadius=0,maxRadius=0)
circles = np.uint16(np.around(circles))
# print(len(circles[0,:]))
img = image.copy()
# for circle in circles[0,:]:
circle = circles[0,:][0]
# draw the outer circle
def heartSeg(file_name):
# img_blur = cv2.GaussianBlur(img_origin, (3,3), 0, 0)
# img_blur = cv2.blur(img_origin, (5,5))
# Input the image
img_origin_name = 'public/data/' + file_name
img_origin = cv2.imread(img_origin_name, 1)
img_origin_singlechannel = cv2.imread(img_origin_name, 0)
# Image smoothing (also reduce noise)
img_medianblur = cv2.medianBlur(img_origin_singlechannel, 5)
# Apply thresholding to the image twice to get the desired grayscale interval (between 50~145)
thresh_val_high = 145
thresh_val_low = 50
ret, thresh = cv2.threshold(img_medianblur, thresh_val_high, 255,
cv2.THRESH_TOZERO_INV)
ret, img_binary = cv2.threshold(thresh, thresh_val_low, 255,
cv2.THRESH_BINARY)
# Apply morphology opening to binary image to eliminate small objects
kernel = np.ones((3, 3), dtype=np.uint8)
img_obj = cv2.morphologyEx(img_binary,
cv2.MORPH_OPEN,
kernel,
iterations=2)
# Find the absolute foreground (i.e. objects) by eroding all objects
fg = cv2.erode(img_obj, kernel, iterations=3)
# Specify heart wall foreground with the condition of object window size
fg_labels_num, fg_labels_img, fg_stats, fg_centroids = cv2.connectedComponentsWithStats(
fg)
muscle_fg_idx = 0
for muscle_fg_idx in range(fg_labels_num):
if fg_stats[muscle_fg_idx,cv2.CC_STAT_WIDTH] > 80 and fg_stats[muscle_fg_idx,cv2.CC_STAT_WIDTH] < 120 \
and fg_stats[muscle_fg_idx,cv2.CC_STAT_HEIGHT] > 100 and fg_stats[muscle_fg_idx,cv2.CC_STAT_HEIGHT] < 140:
break
fg_labels_img[fg_labels_img != muscle_fg_idx] = 0
fg_labels_img[fg_labels_img == muscle_fg_idx] = 1
muscle_fg = fg_labels_img.astype('uint8')
muscle_fg = cv2.erode(muscle_fg, kernel, iterations=2)
# Specify heart wall region with the condition of object window size
bg = img_obj
bg_labels_num, bg_labels_img, bg_stats, bg_centroids = cv2.connectedComponentsWithStats(
bg)
muscle_bg_idx = 0
for muscle_bg_idx in range(bg_labels_num):
if bg_stats[muscle_bg_idx,cv2.CC_STAT_WIDTH] > 120 and bg_stats[muscle_bg_idx,cv2.CC_STAT_WIDTH] < 160 \
and bg_stats[muscle_bg_idx,cv2.CC_STAT_HEIGHT] > 120 and bg_stats[muscle_bg_idx,cv2.CC_STAT_HEIGHT] < 160:
break
bg_labels_img[bg_labels_img != muscle_bg_idx] = 0
bg_labels_img[bg_labels_img == muscle_bg_idx] = 1
muscle_bg = bg_labels_img.astype('uint8')
# Find the absolute background by dilating heart wall region
muscle_bg = cv2.dilate(muscle_bg, kernel, iterations=7)
muscle_fg = muscle_fg.astype('int32')
muscle_bg = muscle_bg.astype('int32')
# The area between heart wall object and background is unknown area TBD with watershed algorithm
muscle_unknown = np.copy(muscle_bg)
muscle_unknown[muscle_fg > 0] = 0
'''
markers:
>1: absolute foreground (heart wall object)
1: absolute background
0: unknown area (TBD with watershed algorithm)
(all 0 will become 1 (bg) or >1 (fg) after watershed algorithm applied)
-1: watershed boundaries (generated with watershed algorithm)
'''
markers = muscle_fg + 1
markers[muscle_unknown > 0] = 0
# Apply watershed algorithm on Gaussian-blurred image (not the origin one, because the noise is massive)
img_gaussblur = cv2.GaussianBlur(img_origin, (3, 3), 0, 0)
markers = cv2.watershed(img_gaussblur, markers)
# Use our function to generate the image where heart wall region is specified
img_out = getWatershedImage(img_origin, markers)
if not os.path.isdir('public/result'):
os.mkdir('public/result')
cv2.imwrite('public/result/result_' + file_name, img_out)
def watershed(self, original):
hsv = cv2.cvtColor(original, cv2.COLOR_BGR2HSV)
for i in range(len(self.lower_fg)):
self.create_window()
self.fine_tune_fg(self.lower_fg[i], self.upper_fg[i], hsv,
original)
for i in range(len(self.lower_bg)):
self.create_window_bg()
self.fine_tune_bg(self.lower_bg[i], self.upper_bg[i], hsv,
original)
if self.fg is None or self.bg is None:
self.fg = []
self.bg = []
return None, None
self.fg = self.fg_array[0]
for j in range(len(self.fg_array)):
if j > 0:
self.fg = cv2.add(self.fg, self.fg_array[j])
self.bg = self.bg_array[0]
for j in range(len(self.bg_array)):
if j > 0:
self.bg = cv2.add(self.bg, self.bg_array[j])
# passa para gray
fg_gray = cv2.cvtColor(self.fg, cv2.COLOR_BGR2GRAY)
bg_gray = cv2.cvtColor(self.bg, cv2.COLOR_BGR2GRAY)
# simplificar Binarias
_, thresh_fg = cv2.threshold(fg_gray, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
_, thresh_bg = cv2.threshold(bg_gray, 0, 255,
cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
# retirar areas comuns que nao interresam
mark = cv2.subtract(thresh_bg, thresh_fg)
# simplificar markers
mark32 = np.int32(cv2.addWeighted(thresh_fg, 1, thresh_bg, 0.5, 0))
markers = mark32 + 1
markers[mark == 255] = 0
# watershed
markers = cv2.watershed(original, markers)
# encontrar contornos
m = cv2.convertScaleAbs(markers)
_, m = cv2.threshold(m, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
cntr, h = cv2.findContours(m, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_SIMPLE)
cv2.drawContours(original, cntr, -1, (255, 255, 255), 2, cv2.LINE_AA)
self.create_window_view()
while 1:
cv2.imshow(self.filename, original)
k = cv2.waitKey(5) & 0xFF
if k == ord("s"):
cv2.destroyAllWindows()
self.fg = ""
self.bg = ""
self.fg_array = []
self.bg_array = []
return original, cntr
def watershed(self):
m = self.markers.copy()
cv.watershed(self.img, m)
overlay = self.colors[np.maximum(m, 0)]
vis = cv.addWeighted(self.img, 0.5, overlay, 0.5, 0.0, dtype=cv.CV_8UC3)
cv.imshow('watershed', vis)
def main():
#Input the images from the file you want analysed for comparison: img = Original img2 = Labelled
cv2.namedWindow("Results")
img = cv2.imread(r'originals/001.jpg')
img2 = cv2.imread(r'labelled/001.jpg')
#Converting the images into grayscale.
rows, cols = img.shape[0], img.shape[1]
img = cv2.resize(img, (int(cols / 2), int(rows / 2)))
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
ret, thresh = cv2.threshold(gray, 0, 255,
cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
# Noise reduction.
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=1)
# Identifying the background area of the overall image.
sure_bg = cv2.dilate(opening, kernel, iterations=1)
# Determining each individual object. (Cells.)
sure_fg = cv2.erode(opening, kernel, iterations=1)
# Remaining area is the border (seperation area) between the cells and the background.
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Label the interior of each individual cell with a unique label.
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Watershed algorithm identifies the boundaries of each cell.
markers[unknown == 255] = 0
markers = cv2.watershed(img, markers)
# Number of markers is the number of cells in the image.
numbers = markers.max()
# Calculate the gray-scale variance for each cell and treat all cells that are greater than the maximum value of the (gray-scale variance * var_threshold) as diseased and marked as green.
var_threshold = 0.2
mess = []
max_var = 0
for i in range(2, numbers + 1):
if gray[markers == i] != []:
tmp = gray[markers == i]
if tmp.size > 50:
pix_var = tmp.var()
mess.append((pix_var, i, tmp.min()))
if pix_var > max_var:
max_var = pix_var
disease = []
for i in range(len(mess)):
if mess[i][0] >= max_var * var_threshold:
disease.append(mess[i][2])
img[markers == mess[i][1]] = [0, 255, 0] #BGR
# Traverse all the cells once again to prevent any leaks.
# Taking the average of the darkest gray values for each diseased cell as described above, treat all cells that have a darker color than the average* pix_threshold as being diseased and marked in red.
pix_threshold = 0.75
for i in range(len(mess)):
if mess[i][2] <= np.array(mess).mean() * pix_threshold:
img[markers == mess[i][1]] = [0, 0, 255] #BGR
# Map all borders in blue for easy viewing.
img[markers == -1] = [255, 0, 0] #BGR
#Show final results.
print('Total Number Of Cells:' + str(numbers))
(r, g, b) = cv2.split(img)
img = cv2.merge([b, g, r])
(r, g, b) = cv2.split(img2)
img2 = cv2.merge([b, g, r])
plt.figure('Result')
plt.subplot(1, 2, 1)
plt.imshow(img2)
plt.subplot(1, 2, 2)
plt.imshow(img)
plt.show()
# Show grey level histogram.
plt.figure('Gray Level Histogram')
arr = gray.flatten()
n, bins, patches = plt.hist(arr,
bins=256,
normed=1,
facecolor='green',
alpha=0.75)
plt.show()
print(unknown)
# Marker labelling
ret, markers = cv.connectedComponents(sure_fg)
print(markers.shape)
# for x in markers:
# print(x)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
print(markers[unknown == 255])
markers[unknown == 255] = 0
new_img = np.zeros((312, 252, 3), np.uint8)
new_img[:] = (255, 255, 255)
cv.imshow('img1', img)
print(markers)
markers = cv.watershed(img, markers)
markers2 = cv.watershed(new_img, markers)
# print(markers)
img[markers == -1] = [255, 0, 0]
new_img[markers == -1] = [255, 0, 0]
cv.imshow('img', img)
cv.imshow('new_img', new_img)
# cv2.waitKey()
while (1):
# cv2.imshow('newimg', newimg)
if cv.waitKey(1) == ord('q'):
break
def __init__(self, org_img, org_mask):
#画像格納
train_data = org_img[:, :, :3].copy() # omit alpha channel
train_data_gray = cv2.cvtColor(train_data, cv2.COLOR_RGB2GRAY)
mask_data = org_mask
mask_data = mask_data.copy()
#ヒストグラム正規化により差が出やすくする
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
equ = clahe.apply(train_data_gray)
#equからdatas作成、maskと被ってないとこ消去、ノイズ消去
datas = []
for i in range(0, 256):
th = Threshold(equ, i)
thd = th.data()
kernel = np.ones((5, 5), np.uint8)
opening = cv2.morphologyEx(thd, cv2.MORPH_OPEN, kernel) #膨張収縮
closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel) #収縮膨張
inverse = cv2.bitwise_not(closing) #色調反転
data = cv2.bitwise_and(inverse, inverse,
mask=mask_data) #maskの輪郭外を消去
datas.append(data)
datas.append(mask_data) #最後にmaskを入れる
#輪郭配列を作る。
all_contours_array = [] #dataごとのcontoursが全て入っている
for data in datas:
contours_array = [
] #contour_array:各画像に対して[フラグ,moment,輪郭]の形で複数入っている
image, contours, hierarchy = cv2.findContours(
data, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
for contour in contours:
curve_flag = flag.no_curve #くびれがなければ0
if curve(contour): #くびれがあれば1
curve_flag = flag.curve
moment = cv2.moments(contour)
#重心
if moment['m00'] != 0:
cx = int(moment['m10'] / moment['m00'])
cy = int(moment['m01'] / moment['m00'])
else:
cx = 0
cy = 0
contours_array.append([curve_flag, [cx, cy], contour])
all_contours_array.append(contours_array)
seeds = [] #watershedアルゴリズムの前景となる輪郭を入れる配列
first_outer_contours = []
#datas256番=mask_dataで輪郭を取得。くびれていなければseedsへ。くびれていたらfirst_outer_contoursへ。
first_contours_array = all_contours_array[256]
for contour_array in first_contours_array:
if contour_array[0] == flag.curve:
first_outer_contours.append(contour_array[2])
elif contour_array[0] == flag.no_curve:
seeds.append(contour_array[2])
#255番からループスタート
Search(255, first_outer_contours, all_contours_array, seeds)
img = np.copy(train_data)
contour_img = cv2.drawContours(img, seeds, -1, (125, 125, 0), 1)
zero = np.zeros(mask_data.shape)
devided = cv2.fillPoly(zero, pts=seeds,
color=(255, 255, 255)) #分割されたやつ
devided = np.uint8(devided)
#前景となる場所を決める
dist_transform = cv2.distanceTransform(devided, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform,
0.1 * dist_transform.max(), 255, 0)
#背景となる場所を決める
kernel = np.ones((3, 3), np.uint8)
sure_bg = cv2.dilate(mask_data, kernel, iterations=3)
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
ret, markers = cv2.connectedComponents(sure_fg)
markers = markers + 1
# 不明領域に0をつけておく
markers[unknown == 255] = 0
#watershedアルゴリズムを行う
markers = cv2.watershed(train_data, markers)
self.__data = markers
img = cv2.imread('C:/Users/Ferhat/Desktop/img3.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
ret, thresh = cv2.threshold(gray, 0, 255,
cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)
# noise removal
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=2)
# sure background area
sure_bg = cv2.dilate(opening, kernel, iterations=5)
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.2 * dist_transform.max(), 255,
50)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
markers = cv2.watershed(img, markers)
img[markers == -1] = [255, 0, 0]
cv2.imwrite('C:/Users/Ferhat/Desktop/img1213.jpg', img)
## find compromise for overlapping areas
if len(polygon) > 0:
# label non-cell areas
background_label = i + 2
labels[labels == 0] = background_label
# label overlapping areas
labels[cell_count > 1] = 0
# apply watershed algorithm to fill in overlap areas. The "image" entry is an array of zeros, so that the
# boundaries extend uniformly without paying attention to the original histology image (otherwise, the
# boundaries will be crooked)
labels = watershed(np.zeros(im.size[::-1] + (3,), dtype=np.uint8), np.array(labels))
# set background pixels back to zero
labels[labels == background_label] = 0
# compute borders between labels, because in some cases, adjacent cells have intermittent overlaps
# that produce interrupted 0 boundaries
borders = mahotas.labeled.borders(labels, mode='ignore')
# add borders to the labels
labels[borders] = 0
if DEBUG:
plt.subplot(224)
plt.imshow(labels)
def processing(data):
print(data)
file_num = int(len(data))
# Reading 3 images to work
img = [cv2.imread(i, cv2.IMREAD_UNCHANGED) for i in data[:file_num]]
try:
print('Original size', img[0].shape)
myshape = img[0].shape
except AttributeError:
print("shape not found")
# --------------------------------
# setting dim of the resize
height = myshape[0]
width = myshape[1]
dim = (width, height)
res_img = []
for i in range(len(img)):
res = cv2.resize(img[i], dim, interpolation=cv2.INTER_LINEAR)
res_img.append(res)
# Checcking the size
try:
print('RESIZED', res_img[1].shape)
except AttributeError:
print("shape not found")
# Visualizing one of the images in the array
#original = res_img[1]
#display_one(original)
# ----------------------------------
# Remove noise
# Using Gaussian Blur
no_noise = []
for i in range(len(res_img)):
#blur = cv2.GaussianBlur(res_img[i], (5, 5), 0)
blur = cv2.bilateralFilter(res_img[i], 5, 100, 100)
no_noise.append(blur)
for i in range(len(res_img)):
original = res_img[i]
image = no_noise[i]
#display(original, image, 'Original', 'Blured')
im = Image.fromarray(image)
myname = data[i]
mynamestring = str(myname[+image_path_length:])
print("no_noise_path", no_noise_path)
print("my_name", mynamestring)
#im.save(f"{no_noise_path}{mynamestring}")
#---------------------------------
# Segmentation
gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
im5 = Image.fromarray(gray)
ret, thresh = cv2.threshold(gray, 0, 255,
cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
ret, thresh1 = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV)
ret, thresh2 = cv2.threshold(gray, 0, 255, cv2.THRESH_OTSU)
ret, thresh3 = cv2.threshold(gray, 0, 255, cv2.THRESH_OTSU)
th4 = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_MEAN_C,
cv2.THRESH_BINARY_INV, 11, 4)
#increasing to 4 keeps features but removes noise
#display(thresh, thresh2, "Their thresh", "My Thresh")
im2 = Image.fromarray(thresh)
# Further noise removal (Morphology)
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(thresh,
cv2.MORPH_OPEN,
kernel,
iterations=2)
im6 = Image.fromarray(th4)
# sure background area
sure_bg = cv2.dilate(opening, kernel, iterations=3)
# Displaying sure background
#display(original, sure_bg, 'Original', 'Sure background')
# Finding sure foreground area
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform,
0.7 * dist_transform.max(), 255, 0)
# Finding unknown region
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(sure_bg, sure_fg)
#Displaying unknown
#display(original, unknown, 'Original', 'Unknown')
#Displaying segmented back ground
#display(original, sure_bg, 'Original', 'Segmented Background')
im3 = Image.fromarray(sure_bg)
# Marker labelling
ret, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
markers = cv2.watershed(image, markers)
image[markers == -1] = [0, 0, 255]
im4 = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
im4 = Image.fromarray(im4)
im.save(f"{no_noise_path}{mynamestring}")
im2.save(f"{segmented_path}{mynamestring}")
im3.save(f"{segmented_background_path}{mynamestring}")
im4.save(f"{markers_path}{mynamestring}")
im5.save(f"{grey_path}{mynamestring}")
im6.save(f"{adapthresh_path}{mynamestring}")
def get_mask_watershed(image, thresh, aprox_seg, debug=False):
"""Returns the mask of an accurate segmentation based on a approximate
segmentation and an aggresive threshold.
Arguments:
image {cv2 image} -- The image where the search will be perfomed
thresh {np 2d binary array} -- An aggresive threshold mask of the organ
aprox_seg {np 2d binary array} -- An aproximate segmentation of the organ
Keyword Arguments:
debug {bool} -- Set True to see the full contours given by watershed (default: {False})
Returns:
np 2d binary array -- An accurate mask of the organ
"""
width, height = image.shape
# Part of mask we're "sure" is the organ
sure_fg = np.zeros((width, height, 1), np.uint8)
# Part of mask we're "sure" isn't the organ
sure_bg = np.zeros((width, height, 1), np.uint8)
for i in range(width):
for j in range(height):
coord = i, j
sure_bg[coord] = 255
# If it is part of an aggresive threshold,
# and part of an approximate contour, it should be the organ
if aprox_seg[coord] and thresh[i][j]:
sure_fg[coord] = 255
# If it isn't part of either, it should be ignored
elif not aprox_seg[coord] and not thresh[i][j]:
sure_bg[coord] = 0
# Parts we'll have to find what they are
unknown = cv2.subtract(sure_bg, sure_fg)
# Get the markers for watershed
_, markers = cv2.connectedComponents(sure_fg)
# Add one to all labels so that sure background is not 0, but 1
markers = markers + 1
# Now, mark the region of unknown with zero
markers[unknown == 255] = 0
# Wathershed only works with rgb images
backtorgb = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
# Watershed returnes markers of a background, and different foregrounds
markers = cv2.watershed(backtorgb, markers)
# In our case:
# 1 - Background
# 2 - The organ we're interested in
# 3... - Other organs or bones
if debug:
marked = backtorgb
marked[markers == -1] = [0, 255, 0]
cv2.imshow('markers', marked)
cv2.waitKey()
# Since we started the search fron the organ, the first
# foreground region will be the organ of interest
mask = np.zeros((width, height), np.uint8)
for i in range(markers.shape[0]):
for j in range(markers.shape[1]):
if markers[i, j] == 2:
mask[i, j] = 1
return mask
def test(args):
data_loader = CTW1500TestLoader(long_size=args.long_size)
test_loader = torch.utils.data.DataLoader(data_loader,
batch_size=1,
shuffle=False,
num_workers=2,
drop_last=True)
submit_path = 'outputs_shape'
if os.path.exists(submit_path):
shutil.rmtree(submit_path)
os.mkdir(submit_path)
# Setup Model
if args.arch == "resnet50":
model = models.resnet50(pretrained=True,
num_classes=7,
scale=args.scale)
elif args.arch == "resnet101":
model = models.resnet101(pretrained=True,
num_classes=7,
scale=args.scale)
elif args.arch == "resnet152":
model = models.resnet152(pretrained=True,
num_classes=7,
scale=args.scale)
for param in model.parameters():
param.requires_grad = False
model = model.cuda()
if args.resume is not None:
if os.path.isfile(args.resume):
print("Loading model and optimizer from checkpoint '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
# model.load_state_dict(checkpoint['state_dict'])
d = collections.OrderedDict()
for key, value in checkpoint['state_dict'].items():
tmp = key[7:]
d[tmp] = value
model.load_state_dict(d)
print("Loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
sys.stdout.flush()
else:
print("No checkpoint found at '{}'".format(args.resume))
sys.stdout.flush()
model.eval()
total_frame = 0.0
total_time = 0.0
for idx, (org_img, img) in enumerate(test_loader):
print('progress: %d / %d' % (idx, len(test_loader)))
sys.stdout.flush()
img = Variable(img.cuda(), volatile=True)
org_img = org_img.numpy().astype('uint8')[0]
text_box = org_img.copy()
torch.cuda.synchronize()
start = time.time()
outputs = model(img)
score = torch.sigmoid(outputs[:, 0, :, :])
outputs = (torch.sign(outputs - args.binary_th) + 1) / 2
text = outputs[:, 0, :, :]
kernels = outputs[:, 0:args.kernel_num, :, :] * text
score = score.data.cpu().numpy()[0].astype(np.float32)
text = text.data.cpu().numpy()[0].astype(np.uint8)
kernels = kernels.data.cpu().numpy()[0].astype(np.uint8)
# # c++ version pse
# # pred = pse(kernels, args.min_kernel_area / (args.scale * args.scale))
# # python version pse
# pred = pypse(kernels, args.min_kernel_area / (args.scale * args.scale))
# # scale = (org_img.shape[0] * 1.0 / pred.shape[0], org_img.shape[1] * 1.0 / pred.shape[1])
# scale = (org_img.shape[1] * 1.0 / pred.shape[1], org_img.shape[0] * 1.0 / pred.shape[0])
# label = pred
# label_num = np.max(label) + 1
# bboxes = []
tmp_marker = kernels[-1, :, :]
# for i in range(args.kernel_num-2, -1, -1):
# sure_fg = tmp_marker
# sure_bg = kernels[i, :, :]
# watershed_source = cv2.cvtColor(sure_bg, cv2.COLOR_GRAY2BGR)
# unknown = cv2.subtract(sure_bg,sure_fg)
# ret, marker = cv2.connectedComponents(sure_fg)
# label_num = np.max(marker)
# marker += 1
# marker[unknown==1] = 0
# marker = cv2.watershed(watershed_source, marker)
# marker[marker==-1] = 1
# marker -= 1
# tmp_marker = np.asarray(marker, np.uint8)
sure_fg = kernels[-1, :, :]
sure_bg = text
watershed_source = cv2.cvtColor(sure_bg, cv2.COLOR_GRAY2BGR)
unknown = cv2.subtract(sure_bg, sure_fg)
ret, marker = cv2.connectedComponents(sure_fg)
label_num = np.max(marker)
marker += 1
marker[unknown == 1] = 0
marker = cv2.watershed(watershed_source, marker)
marker -= 1
label = marker
# label = tmp_marker
# scale = (w / marker.shape[1], h / marker.shape[0])
scale = (org_img.shape[1] * 1.0 / marker.shape[1],
org_img.shape[0] * 1.0 / marker.shape[0])
bboxes = []
for i in range(1, label_num):
points = np.array(np.where(label == i)).transpose((1, 0))[:, ::-1]
if points.shape[0] < args.min_area / (args.scale * args.scale):
continue
score_i = np.mean(score[label == i])
if score_i < args.min_score:
continue
# rect = cv2.minAreaRect(points)1
binary = np.zeros(label.shape, dtype='uint8')
binary[label == i] = 1
# a=cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
# print(a)
contours, _ = cv2.findContours(binary, cv2.RETR_TREE,
cv2.CHAIN_APPROX_SIMPLE)
contour = contours[0]
# epsilon = 0.01 * cv2.arcLength(contour, True)
# bbox = cv2.approxPolyDP(contour, epsilon, True)
bbox = contour
if bbox.shape[0] <= 2:
continue
bbox = bbox * scale
bbox = bbox.astype('int32')
bboxes.append(bbox.reshape(-1))
torch.cuda.synchronize()
end = time.time()
total_frame += 1
total_time += (end - start)
print('fps: %.2f' % (total_frame / total_time))
sys.stdout.flush()
for bbox in bboxes:
cv2.drawContours(text_box, [bbox.reshape(bbox.shape[0] // 2, 2)],
-1, (0, 255, 0), 2)
image_name = data_loader.img_paths[idx].split('/')[-1].split('.')[0]
write_result_as_txt(image_name, bboxes,
'outputs_shape/submit_ctw1500/')
text_box = cv2.resize(text_box, (text.shape[1], text.shape[0]))
debug(idx, data_loader.img_paths, [[text_box]],
'outputs_shape/vis_ctw1500/')