def evaluate_(resS, gt, precision, recall, tpr, fpr):
gtFM = gt
gtFM = cv2.compare(gtFM, 128, cv2.CMP_GT)
gtBM = cv2.bitwise_not(gtFM)
gtF = np.sum(gtFM)
gtB = resS.shape[0] * resS.shape[1] * 255 - gtF
mae = 0.
for i in range(NUMBER_THRESHOLD):
resM = np.zeros(resS.shape)
tpM = np.zeros(resS.shape)
fpM = np.zeros(resS.shape)
resM = cv2.compare(resS, i, cv2.CMP_GT)
tpM = cv2.bitwise_and(resM, gtFM)
fpM = cv2.bitwise_and(resM, gtBM)
tp = np.sum(tpM)
fp = np.sum(fpM)
recall[i] += tp / (gtF + EPS)
total = EPS + tp + fp
precision[i] += (tp + EPS) / total
tpr[i] += (tp + EPS) / (gtF + EPS)
fpr[i] += (fp + EPS) / (gtB + EPS)
np.divide(gtFM, 255.0)
np.divide(resS, 255.0)
resS = cv2.absdiff(gtFM, resS)
mae += np.sum(resS) / (gtFM.shape[0] * gtFM.shape[1])
print(mae)
return mae
def cv2_peak_local_max(img, threshold_relative, threshold_abs):
#max_val = img.max()
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(img)
th = max(max_val * threshold_relative, threshold_abs)
_, mm = cv2.threshold(img, th, max_val, cv2.THRESH_TOZERO)
#max filter
kernel = np.ones((3, 3))
mm_d = cv2.dilate(mm, kernel)
loc_maxima = cv2.compare(mm, mm_d, cv2.CMP_GE)
mm_e = cv2.erode(mm, kernel)
non_plateau = cv2.compare(mm, mm_e, cv2.CMP_GT)
loc_maxima = cv2.bitwise_and(loc_maxima, non_plateau)
_, coords, _ = cv2.findContours(loc_maxima, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)
coords = np.array([x.squeeze()[::-1] for cc in coords for x in cc])
coords = np.array(coords)
#coords = np.array(np.where(loc_maxima>0)).T
#the code above is faster than coords = np.array(np.where(loc_maxima>0)).T
return coords
def hist_compare(image_1, image_2):
hist1 = create_rgb_hist(image_1)
hist2 = create_rgb_hist(image_2)
match1 = cv2.compare(hist1, hist2, cv2.HISTCMP_BHATTACHARYYA) #越小越相似
match2 = cv2.compare(hist1, hist2, cv2.HISTCMP_CORREL) #越大越相似
match3 = cv2.compare(hist1, hist2, cv2.HISTCMP_CHISQR) #越小越相似
print("巴氏距离： %s， 相关性： %s， 卡方： %s" % (match1, match2, match3))
def peer_cost(img):
b,g,r = cv2.split(img)
ret, m1 = cv2.threshold(r, 95, 255, cv2.THRESH_BINARY)
ret, m2 = cv2.threshold(g, 40, 255, cv2.THRESH_BINARY)
ret, m3 = cv2.threshold(b, 20, 255, cv2.THRESH_BINARY)
mmax = cv2.max(r, cv2.max(g, b))
mmin = 255
for tmp in b[0]:
if (tmp != 0 and tmp < mmin):
mmin = tmp
for tmp in g[0]:
if (tmp != 0 and tmp < mmin):
mmin = tmp
for tmp in r[0]:
if (tmp != 0 and tmp < mmin):
mmin = tmp
print mmin
ret, m4 = cv2.threshold(mmax - mmin, 15, 255, cv2.THRESH_BINARY)
ret, m5 = cv2.threshold(cv2.absdiff(r, g), 15, 255, cv2.THRESH_BINARY)
m6 = cv2.compare(r, g, cv2.CMP_GT)
m7 = cv2.compare(r, b, cv2.CMP_GE)
mask = m1 & m2 & m3 & m4 & m5 & m6 & m7
return mask
def labfilter(self, image):
'''
This is a filter based on a method proposed in "Fast and Efficient Method for Fire Detection
Using Image Processing" by Turgay Celik
This method uses the CIE L*a*b* color space and performs 4 bitwise filters
The method returns true at any pixel that satisfies:
L* > Lm* (mean of L* values)
a* > am* (mean of a* values)
b* > bm* (mean of b* values)
b* > a*
'''
cieLab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
L, a, b = cv2.split(cieLab)
Lm = cv2.mean(L)
am = cv2.mean(a)
bm = cv2.mean(b)
R1 = cv2.compare(L, Lm, cv2.CMP_GT)
R2 = cv2.compare(a, am, cv2.CMP_GT)
R3 = cv2.compare(b, bm, cv2.CMP_GT)
R4 = cv2.compare(b, a, cv2.CMP_GT)
R12 = cv2.bitwise_and(R1, R2)
R34 = cv2.bitwise_and(R3, R4)
return cv2.bitwise_and(R12, R34)
def peer(img):
b, g, r = cv2.split(img)
ret, m1 = cv2.threshold(r, 95, 255, cv2.THRESH_BINARY)
ret, m2 = cv2.threshold(g, 30, 255, cv2.THRESH_BINARY)
ret, m3 = cv2.threshold(b, 20, 255, cv2.THRESH_BINARY)
mmax = cv2.max(r, cv2.max(g, b))
mmin = cv2.min(r, cv2.min(g, b))
ret, m4 = cv2.threshold(mmax - mmin, 15, 255, cv2.THRESH_BINARY)
ret, m5 = cv2.threshold(cv2.absdiff(r, g), 15, 255, cv2.THRESH_BINARY)
m6 = cv2.compare(r, g, cv2.CMP_GE)
m7 = cv2.compare(r, b, cv2.CMP_GE)
mask = m1 & m2 & m3 & m6 & m4 & m5 & m7
cv2.imshow("b", b)
cv2.imshow("g", g)
cv2.imshow("r", r)
cv2.imshow('r_thre', m1)
cv2.imshow('g_thre',m2)
cv2.imshow('b_thre',m3)
cv2.imshow('max-min',m4)
cv2.imshow('absdiff',m5)
cv2.imshow('r_g',m6)
cv2.imshow('r_b',m7)
cv2.imshow('res',mask)
return mask
def rgbfilter_white(self, image,image_bg):
rd = 2
rd2 = 100
diff = cv2.subtract(cv2.cvtColor(image,cv2.COLOR_BGR2GRAY),image_bg)
res = cv2.compare(diff,rd,cv2.CMP_GT)
res1 = cv2.compare(diff,rd2,cv2.CMP_LT)
return cv2.bitwise_and(res,res1)
def getLocalMaxima(image, minDistance, seType=None):
"""
Returns local maxima in image.
Based on http://stackoverflow.com/questions/5550290/find-local-maxima-in-grayscale-image-using-opencv/21023493#21023493
The image will be dilated with structure element.
:param image:
:type image: np.ndarray
:param minDistance: size of the structure element
:type minDistance: int
:param seType: structure element type i.e. cv2.MORPH_RECT
:type seType: int
:return: image with local maximas
:rtype: np.ndarray
"""
size = 2 * minDistance + 1
if seType is None:
seType = cv2.MORPH_RECT
kernel = cv2.getStructuringElement(seType, (size, size))
dil = cv2.dilate(image, kernel=kernel)
# no changes in
mask = cv2.compare(image, dil, cmpop=cv2.CMP_EQ)
localMaxiImage = cv2.compare(image, 0, cmpop=cv2.CMP_GT)
localMaxiImage[mask == 0] = 0
# Remove the image borders
localMaxiImage[:minDistance] = 0
localMaxiImage[-minDistance:] = 0
localMaxiImage[:, :minDistance] = 0
localMaxiImage[:, -minDistance:] = 0
return localMaxiImage
def depthBackgroundSubtract(depth, staticMap=None):
if not staticMap:
staticMap = cv2.imread("staticmapCV2.png", cv2.CV_LOAD_IMAGE_UNCHANGED)
nodepthmask = cv2.compare(depth, np.array(10), cv2.CMP_LE)
depth[nodepthmask != 0] = 10000
return cv2.compare(depth, staticMap, cv2.CMP_LT)
def depthBackgroundSubtract(depth, depth_diff_mask, staticMap=None):
if not staticMap:
staticMap = cv2.imread("staticmapCV2.png", cv2.CV_LOAD_IMAGE_UNCHANGED)
np.zeros(depth.shape, dtype="uint16")
nodepthmask = cv2.compare(depth, np.array(10), cv2.CMP_LE)
depth[nodepthmask != 0] = 10000
depth_diff_mask = cv2.compare(depth, staticMap, cv2.CMP_LT)
def feed(self, image):
if self.backImg == None:
self.backImg = image.copy()
self.prevImg = image.copy()
self.height, self.width = image.shape
self.bestRun = np.ones((self.height,self.width), np.uint8)
self.currentRun = np.ones((self.height,self.width), np.uint8)
self.minFixedPixel = int(image.size*self.PERCENTAGE)
#print self.minFixedPixel
if self.showBackImage:
self.createTrackbars()
cv2.imshow("backImage", self.backImg)
return self.backImg
self.checkSettings()
diffImage = cv2.absdiff(self.prevImg,image)
ret, threshold1 = cv2.threshold(diffImage, self.THRESHOLD, 1, cv2.THRESH_BINARY_INV)
ret, threshold255 = cv2.threshold(diffImage, self.THRESHOLD, 255, cv2.THRESH_BINARY_INV)
nonZero = cv2.countNonZero(threshold1)
nonZeroRatio = nonZero / float(image.size)
perfection = self.PERFECTION
if nonZeroRatio < self.PERCENTAGE:
perfection = 5
print perfection
nonChanged = cv2.bitwise_and(self.currentRun, threshold255)
self.currentRun = cv2.add(threshold1,nonChanged)
newBestsMask = cv2.compare(self.currentRun, self.bestRun, cv2.CMP_GE)
oldBestsMask = cv2.compare(self.currentRun, self.bestRun, cv2.CMP_LT)
newBestRuns = cv2.bitwise_and(self.currentRun, self.currentRun, mask = newBestsMask)
oldBestRuns = cv2.bitwise_and(self.bestRun, self.bestRun, mask = oldBestsMask)
self.bestRun = cv2.add(newBestRuns, oldBestRuns)
newBackImgPoints = cv2.bitwise_and(image, image,mask = newBestsMask)
oldBackImgPoints = cv2.bitwise_and(self.backImg, self.backImg, mask = oldBestsMask)
self.backImg = cv2.add(newBackImgPoints, oldBackImgPoints)
stablePoints = cv2.compare(self.bestRun, perfection, cv2.CMP_GT)
unstablePoints = cv2.bitwise_not(stablePoints)
stablePoints = cv2.bitwise_and(stablePoints, perfection)
unstablePoints = cv2.bitwise_and(unstablePoints, self.bestRun)
self.bestRun = cv2.add(stablePoints, unstablePoints)
self.nonZeroPoints = cv2.countNonZero(stablePoints)
self.prevImg = image.copy()
if self.showBackImage:
cv2.imshow("backImage", self.backImg)
return self.backImg
def midThreshold(img, minValue, maxValue, value):
# anything above minValue and below maxValue becomes value
# anything outside of min/maxValue maintains its value
gt = cv2.compare(img, minValue, cv2.CMP_GE)
lt = cv2.compare(img, maxValue, cv2.CMP_LE)
between = cv2.bitwise_and(gt, lt)
newImg = np.zeros(img.shape, dtype=np.uint8)
newImg += img
newImg -= img * (between / 255)
newImg += value * (between / 255)
return newImg
def midThreshold(img, minValue, maxValue, value): # anything above minValue and below maxValue becomes value # anything outside of min/maxValue maintains its value gt = cv2.compare(img, minValue, cv2.CMP_GE) lt = cv2.compare(img, maxValue, cv2.CMP_LE) between = cv2.bitwise_and(gt, lt) newImg = np.zeros(img.shape, dtype=np.uint8) newImg += img newImg -= img*(between/255) newImg += value*(between/255) return newImg
def peer2(img):
b, g, r = cv2.split(img)
ret, m1 = cv2.threshold(r, 220, 255, cv2.THRESH_BINARY)
ret, m2 = cv2.threshold(g, 210, 255, cv2.THRESH_BINARY)
ret, m3 = cv2.threshold(b, 170, 255, cv2.THRESH_BINARY)
ret, m4 = cv2.threshold(cv2.absdiff(r, g), 15, 255, cv2.THRESH_BINARY)
m5 = cv2.compare(r, g, cv2.CMP_GT)
m6 = cv2.compare(r, b, cv2.CMP_GE)
mask = m1 & m2 & m3 & m4 & m5 & m6
return mask
def findIou(refMask, detMask):
diff = cv2.compare(refMask,detMask,cv2.CMP_NE)
detCopy = detMask.copy()
detCopy[np.where(detCopy == [0])] = [128]
intersection = cv2.compare(refMask,detCopy,cv2.CMP_EQ)
numInterSection = cv2.countNonZero(intersection)
numDiff = cv2.countNonZero(diff)
iou = float(numInterSection)/float(numInterSection+numDiff)
return iou
def normalizeImage(self, new_size):
# create a resized version of the image
# with a predefined size
# first, generate a squared version of original
#...check longest to keep aspect ratio....
longest = max(self.img.shape[0], self.img.shape[1])
#...offset....
offset_y = int((longest - self.img.shape[0]) / 2.0)
offset_x = int((longest - self.img.shape[1]) / 2.0)
#...check if padding is required...
if longest < ConnectedComponent.MinScalingSize:
padding = int(
math.ceil((ConnectedComponent.MinScalingSize - longest) / 2.0))
else:
padding = 0
start_y = offset_y + padding
start_x = offset_x + padding
#...the final matrix...
squared = np.zeros((longest + padding * 2, longest + padding * 2))
# generate the squared version ...
squared[start_y:(start_y + self.img.shape[0]),
start_x:(start_x + self.img.shape[1])] = self.img
# now generate the scaled version
#scaled = np.zeros((new_size, new_size))
scaled = cv2.resize(squared, (new_size, new_size))
self.normalized = cv2.compare(scaled, 128, cv2.CMP_GT)
def find_max_blob(self, blobs):
"""
:param blobs:
:return: 面積が最大のblobを求めるかんせうう
"""
i = 1
b = cv2.compare(blobs, i, cv2.CMP_EQ)
x, y, w, h = cv2.boundingRect(b)
max_area = w * h
for j in range(2, len(blobs)):
b = cv2.compare(blobs, j, cv2.CMP_EQ)
x, y, w, h = cv2.boundingRect(b)
if max_area < w * h:
i = j
max_area = w * h
return cv2.compare(blobs, i, cv2.CMP_EQ)
def peer(img):
b, g, r = cv2.split(img)
ret, m1 = cv2.threshold(r, 95, 255, cv2.THRESH_BINARY)
ret, m2 = cv2.threshold(g, 40, 255, cv2.THRESH_BINARY)
ret, m3 = cv2.threshold(b, 20, 255, cv2.THRESH_BINARY)
mmax = cv2.max(r, cv2.max(g, b))
mmin = cv2.min(r, cv2.min(g, b))
ret, m4 = cv2.threshold(mmax - mmin, 15, 255, cv2.THRESH_BINARY)
ret, m5 = cv2.threshold(cv2.absdiff(r, g), 15, 255, cv2.THRESH_BINARY)
m6 = cv2.compare(r, g, cv2.CMP_GT)
m7 = cv2.compare(r, b, cv2.CMP_GE)
mask = m1 & m2 & m3 & m4 & m5 & m6 & m7
return mask
def segmentAndMask(rgb, depth, mask, rgb_final, depth_final, thresh, staticMap = None):
if not staticMap:
staticMap = cv2.imread("staticmapCV2.png", cv2.CV_LOAD_IMAGE_UNCHANGED)
rgb_final = rgb.copy()
depth_final = depth.copy()
# only keep foreground
foregroundMask = np.zeros(depth.shape, dtype="uint8")
depthBackgroundSubtract(depth, foregroundMask)
backgroundMask = np.zeros(depth.shape, dtype="uint8")
cv2.bitwise_not(foregroundMask, backgroundMask)
rgb_final[backgroundMask != 0] = 0
depth_final[backgroundMask != 0] = 0
# mask out the pixels with no depth
noDepthMask = cv2.compare(depth, np.array(0), cv2.CMP_LE)
rgb_final[noDepthMask != 0] = 0
depth_final[noDepthMask != 0] = 0
# mask out pixels specified in mask
rgb_final[mask != 0] = 0
depth_final[mask != 0] = 0
def run(self, img):
#Оцениваем фон
neg_msk, diff = self._run(img)
# Фильтруем
# Для переделки в "неколько градиентов"
# достаточно вызвать self._compute_blured
# с несколькими размерами ядра и
# объединить n, например, - "по или"
neg, d = self._compute_blured(img, self.ng_ksz)
# Т.к. шумы оцениваются "не совсем сверху",
# считаем плотность "шумовых" областей движения.
pos = cv2.compare(neg, 0, cv2.CMP_EQ)
pos = cv2.bitwise_and(pos, 1)
# Единицы там, где может быть "шумовая" составляющая
ns = cv2.bitwise_and(neg_msk, pos)
nsden = np.sum(ns).astype('float') / np.sum(pos)
#Вспышка - резкий рост плотности
flash = self.flash.run(nsden) or self.ifrm.run(nsden)
#Обновление служебной информации
print "Dencity change:", nsden / self.nsden_z
self.nsden_z = nsden
return neg, flash, d
def morphology(img):
# inverts the image to execute easier operations (sum and subtraction)
a, img = cv2.threshold(img, 100, 255, cv2.THRESH_BINARY_INV)
# generates 3 by 3 cross kernel
kernel = cv2.getStructuringElement(cv2.MORPH_CROSS, (3, 3))
# iteration counter
iteration = 0
print("Starting fast thin...")
while 1:
iteration += 1
print("Running iteration", iteration,
"of morphology") # "Running iteration x"
#erosion
last_img = img.copy()
ero = cv2.erode(img, kernel, iterations=1)
#dilation
dil = cv2.dilate(ero, kernel, iterations=1)
# result = original - dilated + eroded
img -= dil
img += ero
# ends loop if result is the same from last iteration
if cv2.compare(img, last_img, cv2.CMP_EQ).all():
break
a, img = cv2.threshold(img, 100, 255,
cv2.THRESH_BINARY_INV) # inverts back the image
return img
def plot_intercontour_hist(image, outer_contour_id, contours, graph, normalized=True):
outer_contour = contours[outer_contour_id]
(x, y, width, height) = cv2.boundingRect(outer_contour)
subimage = get_subimage(image, (x, y), (x + width, y + height))
monochrome = cv2.cvtColor(subimage, cv2.COLOR_BGR2GRAY)
inverted_mask = cv2.compare(monochrome, monochrome, cv2.CMP_EQ)
inner_contours = [contours[int(contour_id)] for contour_id in graph.successors(outer_contour_id)]
for i in range(width):
for j in range(height):
point = (x + i, y + j)
outer_contour_test = cv2.pointPolygonTest(outer_contour, point, 0)
inner_contours_test = -1
for inner_contour in inner_contours:
inner_contour_test = cv2.pointPolygonTest(inner_contour, point, 0)
if inner_contour_test > 0:
inner_contours_test = 1
break
if outer_contour_test >= 0 and inner_contours_test < 0:
inverted_mask[j][i] = 0
mask = cv2.bitwise_not(inverted_mask)
cv.Set(cv.fromarray(subimage), WHITE, cv.fromarray(inverted_mask))
inner_contour_id = len(str(inner_contours))
print 'inner contour id: ', inner_contour_id
image_name = '%d-%s'%(outer_contour_id, inner_contour_id)
#cv2.imshow(image_name, subimage)
#subhists = plot_hist(subimage, mask, image_name)
(subhists, winnames) = plot_hist_hls(subimage, mask, image_name, normalized)
return subhists, subimage, mask, x, y, winnames
def morphology(img):
# Menentukan nilai treshhold
a, img = cv2.threshold(img,100,255,cv2.THRESH_BINARY_INV)
# Mengubah bentuk dan ukuran kernel
kernel = cv2.getStructuringElement(cv2.MORPH_CROSS,(3,3))
# iteration
iteration = 0
print ("Starting fast thin...")
while 1:
iteration += 1
print ("Running iteration",iteration,"of morphology")
#erosion
last_img = img.copy()
ero = cv2.erode(img,kernel,iterations = 1)
#dilation
dil = cv2.dilate(ero,kernel,iterations = 1)
# result
img -= dil
img += ero
# Mengakhiri loop jika hasilnya sama dari iterasi terakhir
if cv2.compare(img, last_img, cv2.CMP_EQ).all():
break
a, img = cv2.threshold(img,100,255,cv2.THRESH_BINARY_INV)
return img
def laplaceEdge(self, dimg, threshold):
edge = None
if dimg is not None:
blur = cv2.GaussianBlur(dimg, (5,5), sigmaX=0)
laplacian = np.abs(cv2.Laplacian(blur, cv2.CV_32F, ksize=3))
edge = cv2.compare(laplacian, threshold, cv2.CMP_GT)
return edge
def get_foreground(img):
"""Divide the foreground && background of the given image"""
h, w = img.shape[:2]
rectangle = (1, 1, w, h)
result = np.zeros(shape=(h, w), dtype=np.uint8)
bgModel = np.zeros((1, 13 * 5))
fgModel = np.zeros((1, 13 * 5))
cv.grabCut(img, result, rectangle, bgModel, fgModel, 5, cv.GC_INIT_WITH_RECT)
#cv.imshow("test", result)
#cv.waitKey(100)
temp1 = np.array(cv.GC_PR_FGD, dtype=np.uint8)
result2 = cv.compare(result, temp1, cv.CMP_EQ)
cv.imshow("test", result2)
cv.waitKey(2000)
#print result2
x, y = img.shape[:2]
background = np.ones(shape=(x, y, 3), dtype=np.uint8)
#background *= 255
foreground = np.ones(shape=(x, y, 3), dtype=np.uint8)
#foreground *= 255
for i in range(0, x):
for j in range(0, y):
if (result2[i, j].all() == 0):
background[i, j] = img[i, j]
else:
foreground[i, j] = img[i, j]
#print background
cv.imshow("test", foreground)
cv.waitKey(2000)
cv.imshow("test1", background)
cv.waitKey(2000)
cv.imwrite("prueba.png", foreground)
return foreground
def segmentation2contours(seg_mask, kernel_size = 5, chain_approx = False):
chain_approx = cv2.CHAIN_APPROX_SIMPLE if chain_approx else cv2.CHAIN_APPROX_NONE
n_labs, seg_mask, stats, centroids = cv2.connectedComponentsWithStats(seg_mask.astype(np.uint8), 4)
kernel = cv2.getStructuringElement( cv2.MORPH_ELLIPSE, (kernel_size, kernel_size))
mm = kernel_size // 2 #offset
cell_contours = []
for ii, (top, left, height, width, _) in enumerate(stats):
if ii == 0:
continue
yl, yr = max(0, left - mm), min(seg_mask.shape[0], left + width + mm)
xl, xr = max(0, top - mm), min(seg_mask.shape[1], top + height + mm)
crop = cv2.compare(seg_mask[yl:yr, xl:xr], ii, cv2.CMP_EQ)
crop = cv2.dilate(crop, kernel)
cnt = cv2.findContours(crop, cv2.RETR_EXTERNAL, chain_approx)[-2]
assert len(cnt) == 1
cnt = cnt[0].squeeze(1) + np.array((xl, yl))[None]
cell_contours.append(cnt)
return cell_contours
def getROI( img, Nblur, Nsobel, Bsepare ):
if len(img.shape)>2:
img_gray = rgb2gray( img, );
else:
img_gray = img;
IM_WIDTH = img_gray.shape[1]
IM_HEIGHT = img_gray.shape[0]
########################################################################
numel = IM_WIDTH * IM_HEIGHT;
NordB = Nblur;
NordS = Nsobel;
if NordB>0:
img_gray = cv2.blur(img_gray,(NordB,NordB))
sobelx = cv2.Sobel(img_gray,cv2.CV_64F,1,0,ksize=NordS)
sobely = cv2.Sobel(img_gray,cv2.CV_64F,0,1,ksize=NordS)
sobel2 = sobelx * sobelx + sobely * sobely
sobel = cv2.compare( sobel2, int((8*np.sum(sobel2)/(numel))), cv2.CMP_GT )
sobel[0:IM_HEIGHT, 0:NordS] = 0
sobel[0:IM_HEIGHT, IM_WIDTH-NordS:IM_WIDTH] = 0
sobel[0:NordS,0:IM_WIDTH] = 0
sobel[IM_HEIGHT-NordS:IM_HEIGHT,0:IM_WIDTH] = 0
sobelf = cv2.bitwise_and( imfill( sobel, 8 ), cv2.bitwise_not(sobel) );
#
sobelf = imclose( sobelf, 'square', NordS )
#
if Bsepare:
mxv = int(round( np.max(cv2.distanceTransform( sobelf, cv2.DIST_L2, 3 ))/2 ))
sobelf = imopen( sobelf, 'square', int(mxv) );
else:
sobelf = imopen( sobelf, 'square', int(15) )
#
# cv2.imshow( "Color", (sobel) );cv2.waitKey()
#
# sobelf = imfill( sobelf, 8 );
return sobelf
def corner_detection():
image_filepath = '../../../data/machine_vision/build.png'
img = cv.imread(image_filepath, cv.IMREAD_GRAYSCALE)
if img is None:
print('Failed to load an image, {}.'.format(image_filepath))
return
dst_32f = cv.cornerHarris(img, blockSize=3, ksize=3, k=0.01)
dilated = cv.dilate(dst_32f, kernel=None, anchor=None, iterations=1)
localMax = cv.compare(dst_32f, dilated, cv.CMP_EQ) # Binary image.
minv, maxv, minl, maxl = cv.minMaxLoc(dst_32f)
_, dst_32f = cv.threshold(dst_32f, 0.01 * maxv, 255, cv.THRESH_BINARY)
cornerMap = (dst_32f * 255).astype(np.int32)
cornerMap[localMax == 0] = 0 # Deletes all modified pixels.
centers = np.where(cornerMap > 0)
#centers = np.vstack(centers.tolist()).T
centers = list(zip(centers[1].tolist(), centers[0].tolist()))
for center in centers:
cv.circle(img, center, 3, (255, 255, 255), 2, cv.LINE_AA)
cv.imshow('Image', img)
cv.imshow('CornerHarris Result', dst_32f)
cv.imshow('Unique Points after Dilatation/CMP/And', cornerMap)
cv.waitKey(0)
cv.destroyAllWindows()
def estimate_perspective_ii(level_map, img2, img3, operator_position):
"""用可部署地块选出从地图坐标到通常视角的透视矩阵。
同名参数含义与estimate_perspective函数一致，返回值也一致，但是透视矩阵是三维的。
"""
height, width = img2.shape[:2]
mask = cv2.compare(cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY),
cv2.cvtColor(img3, cv2.COLOR_BGR2GRAY), cv2.CMP_NE)
perspectives = [
generate_perspective((width, height), level_map, view, True)
for view in range(4)
]
# 寻找最匹配的一种视角。
view = np.argmax([
# 峰值信噪比是一种评价图像质量的客观标准……
cv2.PSNR(
mask,
cv2.inRange(
generate_bullet_time_buildable_mask(
(width, height), level_map, perspective,
operator_position), 16, 255))
for perspective in perspectives
])
return generate_perspective((width, height), level_map, view,
False), perspectives[view]
def ycrcbfilter(self, image):
ycrcb = cv2.cvtColor(image, cv2.COLOR_BGR2YCR_CB)
Y, Cr, Cb = cv2.split(ycrcb)
ycb = cv2.compare(Y, Cb, cv2.CMP_GT)
crcb = cv2.compare(Cr, Cb, cv2.CMP_GT)
A1 = cv2.bitwise_and(ycb, crcb)
ym = cv2.mean(Y)
yym = cv2.compare(Y, ym, cv2.CMP_GT)
A2 = cv2.bitwise_and(A1, yym)
crm = cv2.mean(Cr)
ccrm = cv2.compare(Cr, crm, cv2.CMP_GT)
A3 = cv2.bitwise_and(A2, ccrm)
cbm = cv2.mean(Cb)
cbcbm = cv2.compare(Cb, cbm, cv2.CMP_LT)
return cv2.bitwise_and(A3, cbcbm)
def start():
freeze = False
counter = 0
while True:
try:
t1 = time.time()
frames = pipeline.wait_for_frames()
aligned_frames = align.process(frames)
aligned_depth_frame = aligned_frames.get_depth_frame()
aligned_depth_frame = depth_filter(aligned_depth_frame)
color_frame = aligned_frames.get_color_frame()
color_image = np.asanyarray(color_frame.get_data())
# Apply flood fill
points = pc.calculate(aligned_depth_frame)
vtx = np.ndarray(buffer=points.get_vertices(),
dtype=np.float32,
shape=(d1 * d2, 3))
o3d_pointcloud.points = o3d.utility.Vector3dVector(vtx)
if counter == 0:
vis.create_window(window_name="Open3d", width=400, height=400)
vis.add_geometry(o3d_pointcloud)
else:
vis.update_geometry(o3d_pointcloud)
vis.poll_events()
vis.update_renderer()
x = vtx[f1, :] - vtx[f2, :]
y = vtx[f3, :] - vtx[f4, :]
xyz_norm = normalize_v3(np.cross(x, y))
norm_flat = xyz_norm @ norm
norm_fill = norm_flat # norm_fill[rand_idx] = norm_flat
norm_matrix = np.abs(norm_fill.reshape((d1, d2)))
norm_umatrix = cv2.resize(cv2.UMat(norm_matrix), (d2 * 4, d1 * 4))
bool_matrix = cv2.compare(norm_umatrix, 0.95, cmpop=cv2.CMP_GT)
comps, out, stats, cents = getConnects(bool_matrix, connectivity=4)
sizes = stats.get()[:, 2] # get the area sizes
max_label = np.argmax(sizes[2:comps]) + 2
color_image[out.get() == max_label] = 255
t2 = time.time()
text = f'FPS: {1 / (t2 - t1)}'
font = cv2.FONT_HERSHEY_SIMPLEX
color = (255, 0, 0)
place = (50, 50)
thicc = 2
color_image = cv2.putText(color_image, text, place, font, 1, color,
thicc, cv2.LINE_AA)
cv2.imshow('Color', color_image)
if cv2.waitKey(100) & 0xFF == ord('q'):
break
counter += 1
except:
pass
cv2.destroyAllWindows()
pipeline.stop()
def simularity(image, target):
img = cv2.imread(image, 0)
tar = cv2.imread(target, 0)
img = get_two(img)
tar = get_two(tar)
ret = cv2.compare(img, tar, cv2.CMP_EQ)
ret = np.sum(ret)
return ret
def p_hash(img):
"""OpenCV pHash的Python再实现。"""
dct = cv2.dct(
cv2.cvtColor(
cv2.resize(img, (32, 32), interpolation=cv2.INTER_LINEAR_EXACT),
cv2.COLOR_BGR2GRAY).astype(np.float32))[:8, :8]
dct[0, 0] = 0
return cv2.compare(dct, float(np.mean(dct)), cv2.CMP_GT) // 255
def stencil_cull(segmentations, stencil, cls=StenCode.car, w=1920, h=1080):
mask = np.full((h, w), 0x7, dtype=np.uint8)
stencil_unmask = np.bitwise_and(stencil, mask)
thresh = cv2.compare(stencil_unmask, cls.value, cv2.CMP_EQ)
thresh = (thresh // 255).astype(np.uint32)
# person_thresh = cv2.compare(stencil_unmask, StenCode.person.value, cv2.CMP_EQ)
# person_thresh = (person_thresh // 255).astype(np.uint32)
return segmentations * thresh
def isMovement(img, threshold = 100):
black = np.zeros((img.shape), np.uint8)
out = cv2.compare(img,black,cv2.CMP_NE)
diff = cv2.countNonZero(img)
if diff < threshold:
return False
else:
return True
def edgeBasedRegionBinarization(image):
#works with gray scale image
if len(image.shape) == 3:
gray_scale = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
else:
gray_scale = image
rows = gray_scale.shape[0]
cols = gray_scale.shape[1]
#get Contrast-Limited Adaptive Histogram Equalization
tiles_y = int(rows / 20) #20
tiles_x = int(cols / 20) #20
equalized = AdaptiveEqualizer.adapthisteq(gray_scale, 0.04, tiles_x,
tiles_y)
#now apply Canny edge detection....
edges = cv2.Canny(gray_scale, 10, 50, apertureSize=3)
#dilate the edges...
#....create structuring element...
strel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
#....now dilate edges...
dilate = cv2.dilate(edges, strel)
#now invert...
board = cv2.bitwise_not(dilate)
#now label connected components...
labels, count_labels = scipy.ndimage.measurements.label(board)
#get the sizes of the connected components...
sizes = scipy.ndimage.measurements.sum(board, labels,
range(count_labels + 1))
#filter all connected components below certain threshold...
percent = 0.05
remove_size = (sizes / 255.0) < (rows * cols * percent)
remove_pix = remove_size[labels]
labels[remove_pix] = 0
#it has to be above certain percentage to be background
only_board = cv2.compare(labels, 0, cv2.CMP_GT)
#...dilate...
only_board = cv2.dilate(only_board, strel)
#print("Min " + str(equalized.min()))
#print("Min " + str(equalized.max()))
#cv2.imshow("test", equalized / 255.0)
#cv2.waitKey(0)
#final_content = np.zeros( equalized.shape )
final_content = Binarizer.threshold_content(equalized, only_board, 128)
return final_content
def rgbfilter_gray(self, image, rgbthreshold):
b,g,r = cv2.split(image)
rd = rgbthreshold
min1 = cv2.min(b,g)
min1 = cv2.min(min1,r)
max1 = cv2.max(b,g)
max1 = cv2.max(max1,r)
diff = cv2.absdiff(max1,min1)
res = cv2.compare(diff,rd,cv2.CMP_LT)
return res
def segmentAndMask(depth):
depth_final = depth.copy()
# only keep foreground
foregroundMask = depthBackgroundSubtract(depth)
backgroundMask = np.zeros(depth.shape, dtype="uint8")
backgroundMask = cv2.bitwise_not(foregroundMask)
depth_final[backgroundMask != 0] = 0
# mask out the pixels with no depth
noDepthMask = cv2.compare(depth, np.array(0), cv2.CMP_LE)
depth_final[noDepthMask != 0] = 0
return depth_final
def plot_inner_hist(image, outer_contour_id, contours):
outer_contour = contours[outer_contour_id]
(x, y, width, height) = cv2.boundingRect(outer_contour)
subimage = get_subimage(image, (x, y), (x + width, y + height))
monochrome = cv2.cvtColor(subimage, cv2.COLOR_BGR2GRAY)
inverted_mask = cv2.compare(monochrome, monochrome, cv2.CMP_EQ)
for i in range(width):
for j in range(height):
point = (x + i, y + j)
outer_contour_test = cv2.pointPolygonTest(outer_contour, point, 0)
if outer_contour_test > 0:
inverted_mask[j][i] = 0
mask = cv2.bitwise_not(inverted_mask)
cv.Set(cv.fromarray(subimage), (0, 0, 0), cv.fromarray(inverted_mask))
image_name = '%d'%(outer_contour_id)
#cv2.imshow(image_name, subimage)
#subhists = plot_hist(subimage, mask, image_name)
(subhists, winnames) = plot_hist_hls(subimage, mask, image_name)
return subhists, subimage, mask, inverted_mask, winnames
def replaceText(self, img, text, rotation): # first, eliminates text that is in img contained by the box color = self.avgColor(img) M = cv2.getRotationMatrix2D((self.x, self.y), self.angle, 1) (rows, cols, ch) = img.shape rotatedImg = np.zeros(img.shape, dtype=np.uint8) (left, top, right, bottom) = self.rotatedBounds() cv2.rectangle(rotatedImg, (left, top), (right-5, bottom-5), color, -1) # adds the text we want to display rotatedImg = addText(rotatedImg, self.rotatedBounds(), text, (10, 10, 10), rotation) M2 = cv2.getRotationMatrix2D((self.x, self.y), -self.angle, 1) unrotatedImg = cv2.warpAffine(rotatedImg, M2, (cols, rows)) newImg = np.zeros(img.shape, dtype=np.uint8) # replaces the box on the image with the updated box isntBlack = cv2.compare(unrotatedImg, newImg, cv2.CMP_NE) / 255 #isBlack = cv2.compare(unrotatedImg, newImg, cv2.CMP_EQ) / 255 newImg += img newImg -= img*isntBlack newImg += unrotatedImg*isntBlack return newImg
def test_cudaarithm_arithmetic(self):
npMat1 = np.random.random((128, 128, 3)) - 0.5
npMat2 = np.random.random((128, 128, 3)) - 0.5
cuMat1 = cv.cuda_GpuMat()
cuMat2 = cv.cuda_GpuMat()
cuMat1.upload(npMat1)
cuMat2.upload(npMat2)
self.assertTrue(np.allclose(cv.cuda.add(cuMat1, cuMat2).download(),
cv.add(npMat1, npMat2)))
self.assertTrue(np.allclose(cv.cuda.subtract(cuMat1, cuMat2).download(),
cv.subtract(npMat1, npMat2)))
self.assertTrue(np.allclose(cv.cuda.multiply(cuMat1, cuMat2).download(),
cv.multiply(npMat1, npMat2)))
self.assertTrue(np.allclose(cv.cuda.divide(cuMat1, cuMat2).download(),
cv.divide(npMat1, npMat2)))
self.assertTrue(np.allclose(cv.cuda.absdiff(cuMat1, cuMat2).download(),
cv.absdiff(npMat1, npMat2)))
self.assertTrue(np.allclose(cv.cuda.compare(cuMat1, cuMat2, cv.CMP_GE).download(),
cv.compare(npMat1, npMat2, cv.CMP_GE)))
self.assertTrue(np.allclose(cv.cuda.abs(cuMat1).download(),
np.abs(npMat1)))
self.assertTrue(np.allclose(cv.cuda.sqrt(cv.cuda.sqr(cuMat1)).download(),
cv.cuda.abs(cuMat1).download()))
self.assertTrue(np.allclose(cv.cuda.log(cv.cuda.exp(cuMat1)).download(),
npMat1))
self.assertTrue(np.allclose(cv.cuda.pow(cuMat1, 2).download(),
cv.pow(npMat1, 2)))
def get_percentage_of_active_pixels_in_frame(self, frame):
# process the frame
self.__foreground_image = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
# calculate the difference between current and background frame
difference = cv2.absdiff(self.__background_image,
self.__foreground_image)
# process the difference
# use median blur
blurred = cv2.medianBlur(difference, 3)
# do the thresholding
thresholded = cv2.compare(blurred, 6, cv2.CMP_GT)
# erode and dilate
eroded = cv2.erode(thresholded, self.__structing_element)
dilated = cv2.dilate(eroded, self.__structing_element)
# store the difference image for further usage
self.__difference_thresholded_image = dilated
# count non zero elements
nonzero_pixels = cv2.countNonZero(dilated)
# calculate the number of non-zero pixels
height, width = self.__foreground_image.shape
percentage_of_nonzero_pixels = (nonzero_pixels * 100 / (height * width))
# prepare data for background update
mask_gt = np.greater(self.__background_image, self.__foreground_image)
mask_lt = np.less(self.__background_image, self.__foreground_image)
# update the background
self.__background_image += mask_lt
self.__background_image -= mask_gt
return percentage_of_nonzero_pixels
doThis = 1
if doThis == 1:
# Run for videos
gCAVid = getContactAngle(varsCA8, headersVid)
gCAVid.analyzeVideos()
elif doThis == 2:
# Run for tilt images
gCATilt = getContactAngle(varsImgTilt, headersImg, isVideo=False)
gCATilt.analyzeTiltImages()
elif doThis ==3:
# Run for static images, they pretend they are videos
gCAStat = getContactAngle(varsImgStatic, headersImg)
gCAStat.analyzeStaticImages()
"""
NOTES:
roi has a minimum size: opencv video.write seems to have a minimum image size
FOURCC CODECs
int fourCC_code = CV_FOURCC('M','J','P','G'); // M-JPEG codec (may not
be reliable).
int fourCC_code = CV_FOURCC('P','I','M','1'); // MPEG 1 codec.
int fourCC_code = CV_FOURCC('D','I','V','3'); // MPEG 4.3 codec.
int fourCC_code = CV_FOURCC('I','2','6','3'); // H263I codec.
int fourCC_code = CV_FOURCC('F','L','V','1'); // FLV1 codec.
MASK LOGIC (CMPOP)
cv2.compare(src1, src2, cmpop[, dst])
CMP_EQ src1 is equal to src2.
drone.reset()
drone.speed = 0.1
W, H = 640, 360
rt = 240
stream = cv2.VideoCapture(path)
while running:
try:
buff = stream.grab()
imageyuv = stream.retrieve(buff)
imagergb = cv2.cvtColor(imageyuv[1],cv2.COLOR_BGR2RGB)
imagehsv = cv2.cvtColor(imagergb,cv2.COLOR_RGB2HSV)
#imagegray = cv2.cvtColor(imagergb,cv2.COLOR_RGB2GRAY)
h,s,v = cv2.split(imagehsv)
r,g,b = cv2.split(imagergb)
gb = cv2.compare(g,b,cv2.CMP_GT)
rg = cv2.compare(r,g,cv2.CMP_GT)
rrt = cv2.compare(r,rt,cv2.CMP_GT)
rgb = cv2.compare(rg,gb,cv2.CMP_EQ)
rgbrt = cv2.compare(rgb,rrt,cv2.CMP_EQ)
sat = cv2.compare(s,50,cv2.CMP_LT)
rgbrts = cv2.compare(sat,rgbrt,cv2.CMP_LT)
#cv2.imshow("RGB",imagergb)
cv2.imshow("Threshold",rgbrt)
cv2.imshow("SAT",sat)
if go == True:
change = rgbrt - lastimg
#cv2.imshow("Change",change)
nc = 0
lastimg = rgbrt
go = True
def run(input_handler: InputHandler, camera_name: str, add_to_db: Optional[str]):
Database.initialize()
im0 = input_handler.get_frame(camera_name).image
im_gray0 = cv2.cvtColor(im0, cv2.COLOR_BGR2GRAY)
im_draw = numpy.copy(im0)
tl, br = util.get_rect(im_draw)
cmt = CMT(im_gray0, tl, br, estimate_rotation=False)
identifier = SubjectIdentifier(add_to_db)
structuring_element = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
while True:
try:
# Read image
im = input_handler.get_frame(camera_name).image
im_gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
im_draw = numpy.copy(im)
cmt.process_frame(im_gray)
# Display results
# Draw updated estimate
if cmt.has_result:
cropped_image = im[cmt.tl[1] : cmt.bl[1], cmt.tl[0] : cmt.tr[0]]
difference = cv2.absdiff(im_gray0, im_gray)
blurred = cv2.medianBlur(difference, 3)
display = cv2.compare(blurred, 6, cv2.CMP_GT)
eroded = cv2.erode(display, structuring_element)
dilated = cv2.dilate(eroded, structuring_element)
cropped_mask = dilated[cmt.tl[1] : cmt.bl[1], cmt.tl[0] : cmt.tr[0]]
cropped_mask[cropped_mask == 255] = 1
horizontal_center = cropped_mask.shape[1] // 2
vertical_center = cropped_mask.shape[0] // 2
cv2.ellipse(
cropped_mask,
(horizontal_center, vertical_center),
(horizontal_center, vertical_center),
0,
0,
360,
3,
-1,
)
cv2.rectangle(im_draw, cmt.tl, cmt.br, (255, 0, 0), 4)
subject = identifier.identify(cropped_image, cropped_mask)
cv2.putText(im_draw, subject.name, cmt.tl, cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2, cv2.LINE_AA)
util.draw_keypoints(cmt.tracked_keypoints, im_draw, (255, 255, 255))
util.draw_keypoints(cmt.votes[:, :2], im_draw)
util.draw_keypoints(cmt.outliers[:, :2], im_draw, (0, 0, 255))
cv2.imshow("main", im_draw)
cv2.waitKey(1)
im_gray0 = im_gray
except IndexError:
Database.save_db()
exit(0)
def matching_pixels(img1, img2):
return cv2.countNonZero(cv2.compare(img1, img2, cv2.CMP_EQ))
def compute_contours(data, threshold_value, transform=None, exact=False, smooth_value=0,
contour_method=cv2.CHAIN_APPROX_TC89_L1, zoom=None):
if exact:
thresh_img = cv2.compare(data, threshold_value, cv2.CMP_EQ)
else:
ret, thresh_img = cv2.threshold(data, threshold_value, 255, 0)
if thresh_img.dtype != np.uint8:
thresh_img = thresh_img.astype(np.uint8)
if zoom:
thresh_img = cv2.resize(thresh_img, (thresh_img.shape[0]*zoom, thresh_img.shape[1]*zoom),
interpolation=cv2.INTER_NEAREST)
thresh_img = cv2.GaussianBlur(thresh_img, (3, 3), 0)
ret, thresh_img = cv2.threshold(thresh_img, 99, 255, 0)
else:
cross = np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=np.uint8)
thresh_img = cv2.dilate(thresh_img, cross)
if thresh_img.dtype != np.uint8:
thresh_img = thresh_img.astype(np.uint8)
if contour_method.lower() in ['simple', 'approx']:
if contour_method.lower == 'simple':
method = cv2.CHAIN_APPROX_SIMPLE
else:
method = cv2.CHAIN_APPROX_TC89_L1
thresh_img, contours, order = cv2.findContours(thresh_img, cv2.RETR_LIST, method)
elif contour_method.lower == 'accurate':
try:
contours = measure.find_contours(thresh_img, threshold_value, fully_connected='high')
except AttributeError as e:
logger.error("Need package 'skimage' for accurate contour method")
raise e
else:
logger.error("Unknown contour method: \"{}\"".format(contour_method))
contours = []
need_column_swap = False
del thresh_img
featurelist = []
contour_len = [len(contour) for contour in contours]
for idx, c_len in enumerate(contour_len):
if c_len > 1:
if smooth_value > 0:
contour = cv2.approxPolyDP(contours[idx].astype(np.float32), smooth_value, True)
else:
contour = contours[idx]
l = len(contour)
if l < 2:
continue
contour = np.reshape(contour, [l, 2])
if transform is not None:
# turn into homogeneous coordinates
con_h = np.ones([l, 3])
# Doing the below is much quicker than the normal:
# contour = np.append(contour, np.ones([l, 1]), axis=1)
if need_column_swap:
con_h[:, :-1] = contour[:, [1, 0]]
else:
con_h[:, :-1] = contour
# apply transform from pixel coords into srs
contour = np.dot(con_h, transform).tolist()
else:
if need_column_swap:
contour[:, [0, 1]] = contour[:, [1, 0]]
contour = contour.tolist()
if contour[0] != contour[-1]:
contour.append(contour[0])
poly = contour
featurelist.append([poly])
return featurelist
def dark_or_light_objects_only(self, color='dark'):
if self.params['circular_mask_x'] != 'none':
if self.image_mask is None:
self.image_mask = np.zeros_like(self.imgScaled)
cv2.circle(self.image_mask,(self.params['circular_mask_x'], self.params['circular_mask_y']),int(self.params['circular_mask_r']),[1,1,1],-1)
self.imgScaled = self.image_mask*self.imgScaled
# If there is no background image, grab one, and move on to the next frame
if self.backgroundImage is None:
reset_background(self)
return
if self.reset_background_flag:
reset_background(self)
self.reset_background_flag = False
return
if self.add_image_to_background_flag:
add_image_to_background(self, color)
self.add_image_to_background_flag = False
return
if self.params['backgroundupdate'] != 0:
cv2.accumulateWeighted(np.float32(self.imgScaled), self.backgroundImage, self.params['backgroundupdate']) # this needs to be here, otherwise there's an accumulation of something in the background
if self.params['medianbgupdateinterval'] != 0:
t = rospy.Time.now().secs
if not self.__dict__.has_key('medianbgimages'):
self.medianbgimages = [self.imgScaled]
self.medianbgimages_times = [t]
if t-self.medianbgimages_times[-1] > self.params['medianbgupdateinterval']:
self.medianbgimages.append(self.imgScaled)
self.medianbgimages_times.append(t)
if len(self.medianbgimages) > 3:
self.backgroundImage = copy.copy(np.float32(np.median(self.medianbgimages, axis=0)))
self.medianbgimages.pop(0)
self.medianbgimages_times.pop(0)
print 'reset background with median image'
try:
kernel = self.kernel
except:
kern_d = self.params['morph_open_kernel_size']
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kern_d,kern_d))
self.kernel = kernel
if color == 'dark':
self.threshed = cv2.compare(np.float32(self.imgScaled), self.backgroundImage-self.params['threshold'], cv2.CMP_LT) # CMP_LT is less than
elif color == 'light':
self.threshed = cv2.compare(np.float32(self.imgScaled), self.backgroundImage+self.params['threshold'], cv2.CMP_GT) # CMP_GT is greater than
elif color == 'darkorlight':
#absdiff = cv2.absdiff(np.float32(self.imgScaled), self.backgroundImage)
#retval, self.threshed = cv2.threshold(absdiff, self.params['threshold'], 255, 0)
#self.threshed = np.uint8(self.threshed)
dark = cv2.compare(np.float32(self.imgScaled), self.backgroundImage-self.params['threshold'], cv2.CMP_LT) # CMP_LT is less than
light = cv2.compare(np.float32(self.imgScaled), self.backgroundImage+self.params['threshold'], cv2.CMP_GT) # CMP_GT is greater than
self.threshed = dark+light
convert_to_gray_if_necessary(self)
# noise removal
self.threshed = cv2.morphologyEx(self.threshed,cv2.MORPH_OPEN, kernel, iterations = 1)
# sure background area
#sure_bg = cv2.dilate(opening,kernel,iterations=3)
# Finding sure foreground area
#dist_transform = cv2.distanceTransform(opening,cv.CV_DIST_L2,3)
#dist_transform = dist_transform / (np.max(dist_transform)) * 255
#ret, sure_fg = cv2.threshold(dist_transform,0.2*dist_transform.max(),255,0)
# Finding unknown region
#sure_fg = np.uint8(sure_fg)
#self.threshed = sure_fg
erode_and_dialate(self)
# publish the processed image
c = cv2.cvtColor(np.uint8(self.threshed), cv2.COLOR_GRAY2BGR)
# commented for now, because publishing unthresholded difference
if OPENCV_VERSION == 2: # cv bridge not compatible with open cv 3, at least at this time
img = self.cvbridge.cv2_to_imgmsg(c, 'bgr8') # might need to change to bgr for color cameras
self.pubProcessedImage.publish(img)
extract_and_publish_contours(self)
#Create a mask where green and one other colour only are almost 1:1
r1 = r * 1.0
g1 = g * 1.0
b1 = b * 1.0
#green:blue
gbratio = cv2.divide(b1,g1)
maskgbratio = cv2.inRange(gbratio,0.7,1.3)
#green:red
grratio = cv2.divide(r1,g1)
maskgrratio = cv2.inRange(grratio,0.7,1.3)
#only g/r or g/b should be near 1, if both then it is whitish, which we don't want
# maskgreenish = cv2.bitwise_xor(maskgrratio,maskgbratio)
#Create a mask for pixels that are green dominant
#ADJUST THIS SO THAT IT IS MORE THAN JUST SLIGHTLY GREEN DOMINANT, BUT IS DEFINATELY A GOOD BIT HIGHER THAN THE OTHERS
mask1 = cv2.compare(g,b,cv2.CMP_GT)
mask2 = cv2.compare(g,r,cv2.CMP_GT)
maskgreendominant = cv2.bitwise_and(mask1, mask2)
verygreen = cv2.bitwise_and(maskgreendominant,cv2.bitwise_not(cv2.bitwise_and(maskgrratio,maskgbratio)))
#Combine green dominant and greenish masks
# colourmask = cv2.bitwise_or(verygreen, maskgreenish)
#Create a mask that sets an intensity threshold
total = cv2.add(cv2.add(r,g),b)
intensitymask = cv2.inRange(total,200,765)
#Combine colour and intensity masks
finalmask = cv2.bitwise_and(verygreen, intensitymask)
output = cv2.bitwise_and(image, image, mask = finalmask)
def compare_rms(img_to_classify):
path = "../coin_images/"
rms_classID = 0
comp_classID = 0
absdif_classID = 0
absdif_score = 99999999
comp_score = 0
rms_score = 99999999
degree = 0
cannyx = 200
crop_size = 55
cropped_coin_only = preprocess_img(img_to_classify)
coin_center = find_center_of_coin(cropped_coin_only)
print "Coin only Center of Coin", coin_center
img_to_classify_cropped = center_crop(cropped_coin_only , coin_center, crop_size)
cv.Canny(img_to_classify_cropped,img_to_classify_cropped ,cv.Round((cannyx/2)),cannyx, 3)
cv2.imwrite("cropped_coin.png", cv2array(img_to_classify_cropped))
#sys.exit(-1)
for name in glob.glob(path+'jheads/*.jpg'):
print name
img = cv2.imread(name)
img_cropped = preprocess_img(img)
cv.Canny(img_cropped,img_cropped ,cv.Round((cannyx/2)),cannyx, 3)
cv2.imwrite("cropped_coin2.png", cv2array(img_cropped))
for x in range(360):
rotated_img = rotate_image(img_cropped ,x)
cropped = cv2array(center_crop(rotated_img, coin_center, crop_size))
cv2.imwrite("rotated.png", cropped)
'''
temp_absdif_score = np.sum(cv2.absdiff(cropped,cv2array(img_to_classify_cropped)))
if temp_absdif_score < absdif_score:
absdif_score = temp_absdif_score
absdif_classID = "1"
absdif_degree = x
print "absdiff:", absdif_score," New classID:", absdif_classID, " Degree:", absdif_degree
#temp_comp_score = np.sum(cv2.compare(cropped, cv2array(img_to_classify_cropped), cv2.CMP_EQ))
'''
img_to_classify_cropped_array = cv2array(img_to_classify_cropped)
img_to_classify_cropped_array_reshaped = np.reshape(img_to_classify_cropped_array, (img_to_classify_cropped_array.shape[0],img_to_classify_cropped_array.shape[1]) )
cropped_reshaped = np.reshape(cropped, (cropped.shape[0],cropped.shape[1]) )
temp_comp_score = compute_ssim(cropped_reshaped, img_to_classify_cropped_array_reshaped)
if temp_comp_score > comp_score:
comp_score = temp_comp_score
comp_classID = "1"
comp_degree = x
print "comp_score:", comp_score," New classID:", comp_classID, " Degree:", comp_degree
rotated_img2 = cv2array(rotate_image(array2cv(img_to_classify) , (360-comp_degree)))
cv2.imwrite("comp_degree.png", rotated_img2)
'''
temp_rms_score = compare_images(cropped , cv2array(img_to_classify_cropped))
if temp_rms_score[0] < rms_score:
rms_score = temp_rms_score[0]
rms_classID = "1"
rms_degree = x
print "RMS Score:", rms_score," New classID:", rms_classID, " Degree:", rms_degree
'''
for name in glob.glob(path+'jtails/*.jpg'):
print name
img = cv2.imread(name)
img_cropped = preprocess_img(img)
cv.Canny(img_cropped,img_cropped ,cv.Round((cannyx/2)),cannyx, 3)
cv2.imwrite("cropped_coin2.png", cv2array(img_cropped))
for x in range(360):
rotated_img = rotate_image(img_cropped ,x)
cropped = cv2array(center_crop(rotated_img, coin_center, crop_size))
cv2.imwrite("rotated.png", cropped)
img_to_classify_cropped_array = cv2array(img_to_classify_cropped)
img_to_classify_cropped_array_reshaped = np.reshape(img_to_classify_cropped_array, (img_to_classify_cropped_array.shape[0],img_to_classify_cropped_array.shape[1]) )
cropped_reshaped = np.reshape(cropped, (cropped.shape[0],cropped.shape[1]) )
temp_comp_score = compute_ssim(cropped_reshaped, img_to_classify_cropped_array_reshaped)
if temp_comp_score > comp_score:
comp_score = temp_comp_score
comp_classID = "2"
comp_degree = x
print "comp_score:", comp_score," New classID:", comp_classID, " Degree:", comp_degree
rotated_img2 = cv2array(rotate_image(array2cv(img_to_classify) , (360-comp_degree)))
cv2.imwrite("comp_degree.png", rotated_img2)
for name in glob.glob(path+'oheads/*.jpg'):
print name
img = cv2.imread(name)
img_cropped = preprocess_img(img)
cv.Canny(img_cropped,img_cropped ,cv.Round((cannyx/2)),cannyx, 3)
cv2.imwrite("cropped_coin2.png", cv2array(img_cropped))
for x in range(360):
rotated_img = rotate_image(img_cropped ,x)
cropped = cv2array(center_crop(rotated_img, coin_center, crop_size))
cv2.imwrite("rotated.png", cropped)
temp_comp_score = np.sum(cv2.compare(cropped, cv2array(img_to_classify_cropped), cv2.CMP_EQ))
if temp_comp_score > comp_score:
comp_score = temp_comp_score
comp_classID = "3"
comp_degree = x
print "comp_score:", comp_score," New classID:", comp_classID, " Degree:", comp_degree
rotated_img2 = cv2array(rotate_image(array2cv(img_to_classify) , (360-comp_degree)))
cv2.imwrite("comp_degree.png", rotated_img2)
for name in glob.glob(path+'otails/*.jpg'):
print name
img = cv2.imread(name)
img_cropped = preprocess_img(img)
cv.Canny(img_cropped,img_cropped ,cv.Round((cannyx/2)),cannyx, 3)
cv2.imwrite("cropped_coin2.png", cv2array(img_cropped))
for x in range(360):
rotated_img = rotate_image(img_cropped ,x)
cropped = cv2array(center_crop(rotated_img, coin_center, crop_size))
cv2.imwrite("rotated.png", cropped)
temp_comp_score = np.sum(cv2.compare(cropped, cv2array(img_to_classify_cropped), cv2.CMP_EQ))
if temp_comp_score > comp_score:
comp_score = temp_comp_score
comp_classID = "4"
comp_degree = x
print "comp_score:", comp_score," New classID:", comp_classID, " Degree:", comp_degree
rotated_img2 = cv2array(rotate_image(array2cv(img_to_classify) , (360-comp_degree)))
cv2.imwrite("comp_degree.png", rotated_img2)
#print "Final classID:", classID, " Degree:", degree
print "ClassID:"
print "1 = Heads"
print "2 = Tails"
print "3 = Other Heads"
print "4 = Other Tails"
print "Final comp_score:", comp_score," Comp classID:", comp_classID, " Degree:", comp_degree
def __ne__(self, other):
return cv2.compare(self._ocvimg, imgunwrap(other), cv2.CMP_NE)
