def detect_lines(self, img):
gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
gray1 = cv2.GaussianBlur(gray,(3,3),5)
gray2 = cv2.GaussianBlur(gray,(5,5),5)
gray3 = cv2.GaussianBlur(gray,(15,15),3)
linesL1 = lsd(gray1)
linesL2 = lsd(gray2)
linesL3 = lsd(gray3)
linesL = np.vstack([linesL1, linesL2])
linesL = np.vstack([linesL, linesL3])
linesL = np.unique(linesL, axis=0)
return linesL
def detect_lines(self, arg):
gray = cv2.cvtColor(arg, cv2.COLOR_RGB2GRAY)
lines = lsd(gray, ang_th=60, sigma_scale=3.0)
if DEBUG:
plt.imshow(arg)
for x1, y1, x2, y2, _ in lines:
plt.gca().add_artist(
Line2D((x1, x2), (y1, y2), linewidth=2,
linestyle='dashed'))
dshow('line segments')
imgHeight, imgWidth, _ = arg.shape
y1max = self.parameter['marginTop'] * imgHeight
y2min = self.parameter['marginBottom'] * imgHeight
x1max = self.parameter['marginLeft'] * imgWidth
x2min = self.parameter['marginRight'] * imgWidth
hlines = []
vlines = []
for x1, y1, x2, y2, w in lines:
dx = abs(x1 - x2)
dy = abs(y1 - y2)
# consider line segments near margins and not too short
if dx > 15 and dy / dx < 0.15 and (y1 < y1max or y2 > y2min):
hlines.append([x1, y1, x2, y2, w])
if dy > 15 and dx / dy < 0.15 and (x1 < x1max or x2 > x2min):
vlines.append([x1, y1, x2, y2, w])
return imgHeight, imgWidth, hlines, vlines
def get_hor_and_ver_merged_lines(image_path,line_type):
micro_soft_ocr = ocrbox.result_microsoft_api(image_path)
img = cv2.imread(image_path, cv2.IMREAD_COLOR)
gray_clr_obj = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
lines_arr = lsd(gray_clr_obj)
word_bounding_boxes = get_word_bounding_boxes(micro_soft_ocr)
text_bounding_boxes = get_text_bounding_boxes(micro_soft_ocr)
if line_type != "tilted" :
line_coord_list=convert_lines_df_to_list(lines_arr)
else:
line_coord_list=convert_lines_df_to_list1(lines_arr)
filter_line_coord_list = remove_lines_inside_text_box(line_coord_list,word_bounding_boxes)
horizontal_values,vertical_values,tilted_lines = getting_horizontal_vertical_lines_values_from_lsd(filter_line_coord_list)
vertical_grouping_list,ver_grouping_lines = group_vertical_lines(vertical_values)
final_table_vcoordinates = get_vertical_merging_lines(vertical_grouping_list)
horizontal_grouping_list = group_horizantal_lines(horizontal_values)
final_table_hcoordinates = get_horizantal_merging_lines(horizontal_grouping_list)
return final_table_vcoordinates, final_table_hcoordinates,tilted_lines
# image_path = "/Users/prasanthpotnuru/Downloads/3_table_2/005.png"
# get_hor_and_ver_merged_lines(image_path)
def extract_segment(path):
img = Image.open(path).transpose(Image.FLIP_TOP_BOTTOM)
# img = img.resize((1160, 1636), Image.NEAREST)
gray = np.asarray(img.convert('L'))
threshold = 96
thresholdedData = (gray > threshold) * 255
thresholdedData = morphology_test(thresholdedData)
thresholdedData = removeSmallObject(thresholdedData, minSize=1000)
lines = lsd(thresholdedData)
fig, ax = plt.subplots() # figsize=(15, 15))
# ax.imshow(img, interpolation='nearest', cmap=plt.cm.gray)
# thick_lines = get_thick_lines(lines)
thick_lines = filter_thick_lines(lines, thresholdedData, 2)
# thick_lines = lines
for i in range(thick_lines.shape[0]):
p0, p1 = (int(thick_lines[i, 0]),
int(thick_lines[i, 1])), (int(thick_lines[i, 2]),
int(thick_lines[i, 3]))
width = thick_lines[i, 4]
ax.plot((p0[0], p1[0]), (p0[1], p1[1]), linewidth=width / 2)
ax.axis('image')
ax.set_xticks([])
ax.set_yticks([])
plt.show()
csv_path = path.split('.')[0] + '.csv'
write_csv(csv_path, np.ndarray.tolist(thick_lines))
def detect_lines(self, arg):
Hline = []
Vline = []
img = cv2.imread(arg, cv2.IMREAD_COLOR)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
imgHeight, imgWidth = gray.shape
lines = lsd(gray)
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
# consider those line whise length more than this orbitrary value
if (abs(pt1[0] - pt2[0]) > 45) and (
(int(pt1[1]) < imgHeight * 0.25) or
(int(pt1[1]) > imgHeight * 0.75)):
# make full horizontal line
Hline.append([0, int(pt1[1]), imgWidth, int(pt2[1])])
if (abs(pt1[1] - pt2[1]) > 45) and (
(int(pt1[0]) < imgWidth * 0.4) or
(int(pt1[0]) > imgWidth * 0.6)):
# make full vertical line
Vline.append([int(pt1[0]), 0, int(pt2[0]), imgHeight])
Hline.sort(key=lambda x: (x[1]), reverse=False)
Vline.sort(key=lambda x: (x[0]), reverse=False)
return img, imgHeight, imgWidth, Hline, Vline
def process_segments(self):
"""Run the line segment detector."""
if self._img is None:
raise ValueError('image must be set before running the line segment detector')
img_gray = np.array(self._img_scaled.convert('L'))
lsd_result = lsd(img_gray, scale=self._lsd_scale)
# sort by distance and keep the 9 longest segments
# add a column to store distance
rows, cols = lsd_result.shape
segments_d = np.empty((rows, cols + 1))
segments_d[:, :-1] = lsd_result
# find distance of each line segment
for i in range(segments_d.shape[0]):
x1, y1, x2, y2, *_ = segments_d[i]
segments_d[i, 5] = np.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2) # distance formula
# sort and remove distance column
lsd_sorted = segments_d[segments_d[:, 5].argsort()[::-1]][:, :-1]
# keep only the 9 longest segments
self.segments = lsd_sorted[:9]
self._segments_processed = True
def detect_lsd_lines(image):
image = image.astype('float64')
if np.max(image) <= 1:
image *= 255
width = image.shape[1]
height = image.shape[0]
scale_w = np.maximum(width, height)
scale_h = scale_w
lsd_lines = lsd(image)
lsd_lines[:, 0] -= width/2.0
lsd_lines[:, 1] -= height/2.0
lsd_lines[:, 2] -= width/2.0
lsd_lines[:, 3] -= height/2.0
lsd_lines[:, 0] /= (scale_w/2.0)
lsd_lines[:, 1] /= (scale_h/2.0)
lsd_lines[:, 2] /= (scale_w/2.0)
lsd_lines[:, 3] /= (scale_h/2.0)
lsd_lines[:, 1] *= -1
lsd_lines[:, 3] *= -1
return {'segments': lsd_lines[:, 0:4], 'nfa': lsd_lines[:, 0:4]}
def detect_line_by_lsd(img00):
# 画像をグレースケール化
gray = cv2.cvtColor(img00, cv2.COLOR_BGR2GRAY)
gray = cv2.GaussianBlur(gray, (5, 5), 5)
# detect
linesL = lsd(gray)
return linesL
def get_corners(edged_img):
lines = lsd(edged_img)
# lines is list of list with value of each list item is
# [point1.x, point1.y, point2.x, point2.y, width]
corners = []
if lines is not None:
horizontal_lines_canvas = np.zeros(edged_img.shape, dtype=np.uint8)
vertical_lines_canvas = np.zeros(edged_img.shape, dtype=np.uint8)
for line in lines:
x1, y1, x2, y2, _ = line
x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
if abs(x2 - x1) > abs(y2 - y1):
(x1, y1), (x2, y2) = sorted(((x1, y1), (x2, y2)), key=lambda pt: pt[0])
cv2.line(horizontal_lines_canvas, (max(x1 - 5, 0), y1), (min(x2 + 5, edged_img.shape[1] - 1), y2),
255, 2)
else:
(x1, y1), (x2, y2) = sorted(((x1, y1), (x2, y2)), key=lambda pt: pt[1])
cv2.line(vertical_lines_canvas, (x1, max(y1 - 5, 0)), (x2, min(y2 + 5, edged_img.shape[0] - 1)),
255, 2)
lines = []
# find the horizontal lines (connected-components -> bounding boxes -> final lines)
(contours, hierarchy) = cv2.findContours(horizontal_lines_canvas, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
contours = sorted(contours, key=lambda c: cv2.arcLength(c, True), reverse=True)[:5]
horizontal_lines_canvas = np.zeros(edged_img.shape, dtype=np.uint8)
for contour in contours:
contour = contour.reshape((contour.shape[0], contour.shape[2]))
min_x = np.min(contour[:, 0]) + 2
max_x = np.max(contour[:, 0]) - 2
left_y = int(np.average(contour[contour[:, 0] == min_x][:, 1]))
right_y = int(np.average(contour[contour[:, 0] == max_x][:, 1]))
lines.append((min_x, left_y, max_x, right_y))
cv2.line(horizontal_lines_canvas, (min_x, left_y), (max_x, right_y), 255, 1)
corners.append((min_x, left_y))
corners.append((max_x, right_y))
# find the vertical lines (connected-components -> bounding boxes -> final lines)
(contours, hierarchy) = cv2.findContours(vertical_lines_canvas, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
contours = sorted(contours, key=lambda c: cv2.arcLength(c, True), reverse=True)[:5]
vertical_lines_canvas = np.zeros(edged_img.shape, dtype=np.uint8)
for contour in contours:
contour = contour.reshape((contour.shape[0], contour.shape[2]))
min_y = np.amin(contour[:, 1]) + 2
max_y = np.amax(contour[:, 1]) - 2
top_x = int(np.average(contour[contour[:, 1] == min_y][:, 0]))
bottom_x = int(np.average(contour[contour[:, 1] == max_y][:, 0]))
lines.append((top_x, min_y, bottom_x, max_y))
cv2.line(vertical_lines_canvas, (top_x, min_y), (bottom_x, max_y), 255, 1)
corners.append((top_x, min_y))
corners.append((bottom_x, max_y))
if corners:
corners = filter_corner(corners) # calling from helpers file
return corners
def return_distribution_length(img):
lines = lsd(img)
length_segments = np.zeros(lines.shape[0])
for i in range(lines.shape[0]):
length_segments[i] = length_segment(lines[i])
length_segments = np.array(length_segments + 1, dtype=np.int)
hist, bin_edges = np.histogram(length_segments,
bins=int(length_segments.max()) - 1,
range=(1, length_segments.max()))
return hist, bin_edges
def drawLineSegments(self, visualize=False):
lines = lsd(self.image)
lines = self.removeFromBoxOfDoom(lines)
lines = list(
filter(lambda line: self.lineThresh(line, self.thresholdLine),
lines))
# # lineFilter
if visualize:
drawLines(lines)
return lines
def findLSDCand(im):
"""
Use LSD algorithm to extract candidate line segment
Args:
im: input image with shape [h, w, 3]
Returns:
cand: a list of all candidate lines, endpoint format, [x1, x2, y1, y2, r], r is not used in this project
"""
gray = cv2.cvtColor(im, cv2.COLOR_BGR2GRAY)
points = lsd(gray)
return points
def lsd_alg(color_image,
line_width=0,
fuse=False,
dTheta=2 / 360 * np.pi * 2,
dRho=2,
maxL=4):
"""
LSD algoritm for line segment detection
:param color_image: [np.array] The input image in BGR mode
line_width: override line width during drawing result
fuse: [bool] Whether to fuse together the segments detected
dTheta: [float] The max difference in theta between two segments to be fused together
dRho: [float] The max difference in rho between two segments to be fused together
maxL: [int] The maximal number of lines fused together. Set to 0 for no limit.
:return: lines: [np.array] (nb_lines, 4) each element of the array corresponds to [pt1_x, pt1_y, pt2_x, pt2_y] in DOUBLES
result_lines: [np.array] BGR images with red lines representing LSD result
result_points: [np.array] BGR images with red dots representing LSD result
"""
gray_image = cv2.cvtColor(color_image, cv2.COLOR_BGR2GRAY)
result_lines = np.zeros(
gray_image.shape +
(3, )) # Black RGB image with same height/width than gray-image
result_points = np.zeros(gray_image.shape + (3, ))
lines = lsd(gray_image) # python script calling the C++ so library
# Fuse lines if asked
if fuse:
lines = lines.reshape((lines.shape[0], 1, lines.shape[1]))
lines = sd.fuseCloseSegment(lines, dTheta, dRho, maxL)
lines = lines.reshape((lines.shape[0], lines.shape[2]))
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
if line_width == 0:
width = lines[i, 4]
else:
width = line_width * 2
cv2.line(result_lines, pt1, pt2, (255, 255, 255),
int(np.ceil(width / 2)))
if 0 <= pt1[0] < gray_image.shape[1] and 0 <= pt1[1] < gray_image.shape[
0]: # Some coordinates returned can be out of bounds
result_points[pt1[1]][pt1[0]][
2] = 255 # Add a red pixel for each end-point of a line
if 0 <= pt2[0] < gray_image.shape[1] and 0 <= pt2[
1] < gray_image.shape[0]:
result_points[pt2[1]][pt2[0]][
2] = 255 # Add a red pixel for each end-point of a line
return lines[:, :
4], result_lines, result_points # Lines over a Black background
def LineSegmentedDetector(image, gray):
resultL = image
#image_copy = image.copy()
linesL = lsd(gray)
#print("lines_lsd:", len(linesL))
for line in linesL:
x1, y1, x2, y2 = map(int, line[:4])
resultL = cv2.line(image, (x1, y1), (x2, y2), (0, 0, 255), 1)
#if (x2-x1)**2 + (y2-y1)**2 > 1000:
# 赤線を引く
# cv2.line(resultL, (x1,y1), (x2,y2), (0,0,255), 3)
return resultL
def _get_horizon_lines(bgr_img):
"""
Given BGR image,
Returns list of green lines (x1,y1,x2,y2,w).
"""
# extract green color into green_mask
# https://stackoverflow.com/questions/47483951/how-to-define-a-threshold-value-to-detect-only-green-colour-objects-in-an-image
img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2HLS)
lower = np.uint8([40, 40, 40])
upper = np.uint8([70, 255, 255])
green_mask = cv2.inRange(img, lower, upper)
return lsd(green_mask)
def detect_line(file_path):
file_path_without_ext, ext = os.path.splitext(file_path)
src = cv2.imread(file_path, cv2.IMREAD_COLOR)
gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
lines = lsd(gray)
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
dist = distance.euclidean(pt1, pt2)
ic(dist, pt1, pt2)
if dist < 15:
continue
width = lines[i, 4]
cv2.line(src, pt1, pt2, (0, 0, 255), int(np.ceil(width / 2)))
out_file_path = file_path_without_ext + '_line.jpg'
ic(out_file_path)
cv2.imwrite(out_file_path, src)
def lineArt(self, image):
fullName = image
src = cv.imread(image, cv.IMREAD_COLOR)
gray = cv.cvtColor(src, cv.COLOR_BGR2GRAY)
lines = lsd(gray)
img = np.zeros([1080,1200,3],dtype=np.uint8)
img.fill(255)
os.mkdir('./lineart')
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
width = lines[i, 4]
cv.line(img, pt1, pt2, (0, 0, 0), int(np.ceil(width / 2)))
if i%10==0 or i==lines.shape[0]-1:
cv.imshow('Step', img)
cv.waitKey(0)
cv.destroyAllWindows()
cv.imwrite('./gallery/new' + i + '.jpg', img)
def constant_width(img, tau=20, thres=.5, compute_param=False):
"""
:param img:
:param tau:
:param thres:
:param compute_param:
:return:
"""
lines = lsd(img)
if compute_param:
thres_values = np.array([.1 * n for n in range(10)])
NFA_values = np.zeros(10)
for i in range(10):
NFA_values[i] = return_NFA(img, tau, thres_values[i])
best_index = np.argmin((NFA_values - 1)**2)
threshold = thres_values[best_index]
print("threshold = ", threshold)
else:
threshold = thres
# test parallel segments
print("\n**** begin parallel test *****")
index_segments = compute_parallel_segments(lines, tau=tau)
# test gradient
print("\n**** begin gradient test *****")
index_segments = return_opposed_gradient(lines, index_segments)
# test aligned segments
print("\n**** begin aligned test *****")
index_segments = return_aligned_segments(lines,
index_segments,
threshold=threshold)
# test proximity
print("\n**** begin proximity test *****")
index_segments = return_proximity_segments(lines, index_segments)
print(f"\n**** NFA= {return_NFA(img, tau, threshold)} *****")
return index_segments
def line_detection(image):
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
lines = lsd(gray)
p_lines = []
image_one = np.ones(np.shape(image)) * 255
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
width = np.sqrt((lines[i, 0] - lines[i, 2])**2 +
(lines[i, 1] - lines[i, 3])**2)
angle = abs(lines[i, 0] - lines[i, 2]) / abs(lines[i, 1] - lines[i, 3])
if width > 10 and (angle < 1):
p_lines.append(
[lines[i, 0], lines[i, 1], lines[i, 2], lines[i, 3]])
cv2.line(image_one, pt1, pt2, (0, 0, 255), int(2))
cv2.imwrite('test_lines.jpg', image_one)
return p_lines
def frame_processing(self, frame):
"""
lsd line extraction
"""
use_lines = []
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
lines = lsd(gray) ################## 引用pylsd 包的部分r
# print(lines)
try:
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
width = lines[i, 4]
if( (eucldist_vectorized(pt1, pt2) < self.cam_LENTHRESHOLD) | (width < self.cam_WIDTHTHRESHOLD) ):
continue
use_lines.append(lines[i])
final_frame = cv2.line(frame, pt1, pt2, (0, 0, 255), int(np.ceil(width / 2)))
return final_frame, use_lines
except :
return frame, lines
def line_detect(src):
'''
detect lines in a given image
Parameters:
fpath - the path of your image
Return:
properties of detected lines
'''
gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
lines = lsd.lsd(gray)
# drop the last column
lines = np.delete(lines, -1, 1)
'''
(lines[i, 0], lines[i, 1]): starting point of the line
(lines[i, 2], lines[i, 3]): ending point of the line
'''
return lines
def detect_vps(frame_gray,
prms,
frame_to_draw=None,
points_staright_old=[],
points_twisted_old=[]):
lines = lsd.lsd(np.array(frame_gray, np.float32))
lines = lines[:, 0:4]
for i in range(lines.shape[0]):
pt1 = (int(lines[i, 0]), int(lines[i, 1]))
pt2 = (int(lines[i, 2]), int(lines[i, 3]))
if (not frame_to_draw is None):
cv.line(frame_to_draw, pt1, pt2, (0, 0, 255), 1)
denoised_lanes = denoise_lanes(lines, prms)
if (len(denoised_lanes) > 0):
points_staright, points_twisted = convert_to_PClines(
denoised_lanes, prms)
else:
points_staright, points_twisted = [], []
detections_straight, m1, b1 = find_detections(points_staright, prms)
detections_twisted, m2, b2 = find_detections(points_twisted, prms)
if (len(detections_straight) == 0 and len(detections_twisted) == 0):
return []
# gather initial vanishing point detections
mvp_all, NFAs = read_detections_as_vps(detections_straight, m1, b1,
detections_twisted, m2, b2, prms)
# refine detections, this returns 2 x ? array ? is the vps left after refining
mvp_all = refine_detections(mvp_all, lines, prms)
mvp_all, NFAs = remove_dublicates(mvp_all, NFAs, prms)
for i in range(len(mvp_all[0])):
p1 = np.int32(mvp_all[:, i])
cv.circle(frame_to_draw, tuple(p1), 5, (0, 255, 0), 3)
return mvp_all.T # return as n x2 shape where n is the number of vps
def estimateFocalLength(image):
from pylsd.lsd import lsd
height = image.shape[0]
width = image.shape[1]
lines = lsd(image.mean(2))
lineImage = image.copy()
for line in lines:
cv2.line(lineImage, (int(line[0]), int(line[1])),
(int(line[2]), int(line[3])), (0, 0, 255),
int(np.ceil(line[4] / 2)))
continue
#cv2.imwrite('test/lines.png', lineImage)
numVPs = 3
VPs, VPLines, remainingLines = calcVanishingPoints(lines, numVPs=numVPs)
#focalLength = (np.sqrt(np.linalg.norm(np.cross(VPs[0], VPs[1]))) + np.sqrt(np.linalg.norm(np.cross(VPs[0], VPs[2]))) + np.sqrt(np.linalg.norm(np.cross(VPs[1], VPs[2])))) / 3
focalLength = (np.sqrt(np.abs(np.dot(VPs[0], VPs[1]))) +
np.sqrt(np.abs(np.dot(VPs[0], VPs[2]))) +
np.sqrt(np.abs(np.dot(VPs[1], VPs[2])))) / 3
return focalLength
def drawLineSegments(file):
global image
rows, cols = image.shape
# gray = np.asarray(image.convert('L'))
# gray = cv2.cvtColor(image,cv2.COLOR_RGB2GRAY)
lines = lsd(image)
# draw = ImageDraw.Draw(image)
if lines is not None:
for line in lines:
# print(line)
pt1 = [int(line[0]), int(line[1])]
pt2 = [int(line[2]), int(line[3])]
# width = lines[i, 4]
dx = int(line[2]) - int(line[0])
dy = int(line[3]) - int(line[1])
dx = dx
dy = dy
length = np.sqrt(dx * dx + dy * dy)
if length > 50:
print(length)
cv2.line(image, tuple(pt1), tuple(pt2), (0, 0, 255), 4)
cv2.imshow("img", image)
def LineDetect(image, thLength):
if image.shape[2] == 1:
grayImage = image
else:
grayImage = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
imageLSD = np.copy(grayImage)
# line segments, [pt1[0], pt1[1], pt2[0], pt2[1], width]
linesLSD = lsd(imageLSD)
del imageLSD
# choose line segments whose length is less than thLength
lineSegs = []
for line in linesLSD:
x1 = line[0]
y1 = line[1]
x2 = line[2]
y2 = line[3]
length = np.sqrt((x1 - x2)**2 + (y1 - y2)**2)
if length > thLength:
lineSegs.append([x1, y1, x2, y2])
return lineSegs
#!/usr/bin/env python # -*- coding: utf-8 -*- # @Date : 2015-12-19 02:09:53 # @Author : Gefu Tang ([email protected]) # @Link : https://github.com/primetang/pylsd # @Version : 0.0.1 import cv2 import numpy as np import os from pylsd.lsd import lsd fullName = 'car.jpg' folder, imgName = os.path.split(fullName) src = cv2.imread(fullName, cv2.IMREAD_COLOR) gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY) lines = lsd(gray) for i in range(lines.shape[0]): pt1 = (int(lines[i, 0]), int(lines[i, 1])) pt2 = (int(lines[i, 2]), int(lines[i, 3])) width = lines[i, 4] cv2.line(src, pt1, pt2, (0, 0, 255), int(np.ceil(width / 2))) cv2.imwrite(os.path.join(folder, 'cv2_' + imgName.split('.')[0] + '.jpg'), src)
def V(im, width, height, focal, params, tmp_count):
# fix random seed
np.random.seed(1)
# principal point is assumed at image center
u0 = width / 2
v0 = height / 2
# line segment (LS) extraction based on the LSD algorithm
# Reference https://github.com/primetang/pylsd
gray = np.asarray(im.convert('L'))
lines = lsd(gray, tmp_count)
lsd_output = lines[:, :4]
# import scipy.io as sio
# mat_lsd = sio.loadmat('lsd.mat')
# lsd_output = mat_lsd['lsd']
# plot the results
if plots.lsd:
im1 = im.copy()
cmap = plt.cm.hsv(np.linspace(0, 1, lsd_output.shape[0]))[:, :3]
draw = ImageDraw.Draw(im1)
for j in range(lsd_output.shape[0]):
pt1 = (lsd_output[j, 0], lsd_output[j, 1])
pt2 = (lsd_output[j, 2], lsd_output[j, 3])
draw.line((pt1, pt2),
fill=tuple((cmap[j] * 255).astype(int)),
width=2)
#im1.show()
im1.save('tmp/tmp1.jpg')
# LS filtering
thres_aligned = max(width, height) / 128.
length_t = np.sqrt(width + height) / 1.71
ls = ls_filter(thres_aligned, length_t, lsd_output)
ls_homo = normalize.normalize(ls, width, height, focal)
# plot the results
if plots.ls_filter:
im2 = im.copy()
cmap = plt.cm.hsv(np.linspace(0, 1, ls.shape[0]))[:, :3]
draw = ImageDraw.Draw(im2)
for j in range(ls.shape[0]):
pt1 = (ls[j, 0], ls[j, 1])
pt2 = (ls[j, 2], ls[j, 3])
draw.line((pt1, pt2),
fill=tuple((cmap[j] * 255).astype(int)),
width=4)
#im2.show()
im2.save('tmp/tmp2.jpg')
# ZL and zenith rough predictions
# prediction of the zenith line
dist_max = width / 8
zl = zl_predict(lsd_output, dist_max, u0, v0, width, height, params)
zl_homo = []
z_homo_cand = []
z_group_cand = []
for i in range(len(zl)):
zl_homo.append(
normalize.normalize(np.array([[zl[i], 0, u0, v0]]), width, height,
focal))
[tmp_z_homo_cand,
tmp_z_group_cand] = z_predict(ls_homo, zl_homo[i], params, 0)
z_homo_cand.append(tmp_z_homo_cand)
z_group_cand.append(tmp_z_group_cand)
# plot the results
if plots.zl:
im3 = im.copy()
cmap = plt.cm.hsv(np.linspace(0, 1, len(z_homo_cand)))[:, :3]
draw = ImageDraw.Draw(im3)
for j in range(len(z_homo_cand)):
# tmp = np.array([-0.0578, -0.9965, 0.0597])
# z = unnormalize.unnormalize(tmp, width, height, focal, 0)
z = unnormalize.unnormalize(z_homo_cand[j], width, height, focal,
0)
pt1 = (width / 2, height / 2)
pt2 = (z[0], z[1])
draw.line((pt1, pt2),
fill=tuple((cmap[j] * 255).astype(int)),
width=4)
# im3.show()
im3.save('tmp/tmp3.jpg')
# choose the best zenith candidate based on the relevance of the predicted HLs
best_z_cand = 0
best_z_score = 0
for i in range(len(zl_homo)):
# HL prediction
[modes_homo, _, _, _, _] = hl_predict(lsd_output, z_homo_cand[i], u0,
v0, width, height, focal, params)
# HL scoring (for performance optimization, each zenith candidate is assessed based only on the meaningful
# HLs (no sampling is performed at that step))
[_, results] = hl_score(modes_homo, ls_homo, z_homo_cand[i], params)
# keep the zenith candidate with highest score
if results["score"] > best_z_score:
best_z_cand = i
best_z_score = results["score"]
# zenith refinement (based on Zhang et al. method)
[z_homo_cand[best_z_cand],
z_group_cand[best_z_cand]] = z_predict(ls_homo, zl_homo[best_z_cand],
params, 1)
# HL prediction
[modes_homo, modes_offset, modes_left, modes_right,
H] = hl_predict(lsd_output, z_homo_cand[best_z_cand], u0, v0, width,
height, focal, params)
# HL sampling
[samp_homo, samp_left,
samp_right] = hl_sample(z_homo_cand[best_z_cand], modes_homo,
modes_offset, modes_left, modes_right, H, u0, v0,
width, height, focal, params)
# plot the results
if plots.hl_samples:
if plots.hl_samples:
im4 = im3.copy()
else:
im4 = im.copy()
cmap = plt.cm.hsv(np.linspace(0, 1, len(zl_homo)))[:, :3]
draw = ImageDraw.Draw(im4)
for j in range(samp_homo.shape[1]):
pt1 = (0, samp_left[j])
pt2 = (width, samp_right[j])
draw.line((pt1, pt2),
fill=tuple((cmap[i] * 255).astype(int)),
width=1)
# im4.show()
im4.save('tmp/tmp4.jpg')
if plots.hl_modes:
if plots.hl_samples and plots.hl_samples:
im5 = im4.copy()
else:
im5 = im.copy()
draw = ImageDraw.Draw(im5)
for j in range(modes_homo.shape[1]):
if H[j] > 0:
pt1 = (0, modes_left[j])
pt2 = (width, modes_right[j])
draw.line((pt1, pt2), fill=tuple([0, 0, 255]), width=4)
# im5.show()
im5.save('tmp/tmp5.jpg')
# HL scoring
# import scipy.io as sio
# tmp_samp_homo = sio.loadmat('samp_homo.mat')
# samp_homo = tmp_samp_homo['samp_homo']
[hl_homo, results] = hl_score(samp_homo, ls_homo, z_homo_cand[best_z_cand],
params)
# import scipy.io as sio
# tmp_hl_homo = sio.loadmat('hl_homo.mat')
# hl_homo = tmp_hl_homo['hl_homo']
# hl_homo = np.squeeze(hl_homo)
hl = unnormalize.unnormalize(hl_homo, width, height, focal, 1)
hvps = unnormalize.unnormalize(results["hvp_homo"], width, height, focal,
0)
hvp_groups = results["hvp_groups"]
z = unnormalize.unnormalize(z_homo_cand[best_z_cand], width, height, focal,
0)
z_group = z_group_cand[best_z_cand]
if params.return_z_homo:
return hl, hvps, hvp_groups, z, z_group, ls, z_homo_cand[
best_z_cand], results["hvp_homo"], ls_homo
#print(0)
return hl, hvps, hvp_groups, z, z_group, ls
img00 = imgcol.copy()
#img00 = cv2.resize(img00,(int(img00.shape[1]/5),int(img00.shape[0]/5)))
gray = cv2.cvtColor(img00, cv2.COLOR_BGR2GRAY)
gray = cv2.GaussianBlur(gray, (5, 5), 5)
t1 = time.time()
edges = cv2.Canny(gray, 50, 150, apertureSize=3)
linesH = cv2.HoughLinesP(edges,
rho=1,
theta=np.pi / 360,
threshold=50,
minLineLength=50,
maxLineGap=10)
t2 = time.time()
linesL = lsd(gray)
t3 = time.time()
img2 = img00.copy()
for line in linesH:
x1, y1, x2, y2 = line[0]
# 赤線を引く
img2 = cv2.line(img2, (x1, y1), (x2, y2), (0, 0, 255), 3)
cv2.imwrite('samp_hagh.jpg', img2)
img3 = img00.copy()
img4 = img00.copy()
for line in linesL:
x1, y1, x2, y2 = map(int, line[:4])
img3 = cv2.line(img3, (x1, y1), (x2, y2), (0, 0, 255), 3)
import cv2
from pylsd.lsd import lsd
import numpy as np
img = cv2.imread('./../sample_image.jpg')
W = 200
w_ratio = 200 / img.shape[1]
img = cv2.resize(img, (200, int(w_ratio * img.shape[0])))
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
gray1 = cv2.GaussianBlur(gray, (3, 3), 5)
gray2 = cv2.GaussianBlur(gray, (5, 5), 5)
gray3 = cv2.GaussianBlur(gray, (15, 15), 3)
#gray1 = cv2.bitwise_not(gray1)
linesL1 = lsd(gray1)
linesL2 = lsd(gray2)
linesL3 = lsd(gray3)
linesL = np.vstack([linesL1, linesL2])
linesL = np.vstack([linesL, linesL3])
#linesL.extend(linesL3)
for line in linesL:
x1, y1, x2, y2 = map(int, line[:4])
dx = x2 - x1
dy = y2 - y1
if (dx)**2 + (dy)**2 > 5000:
#img = cv2.line(img, (x1,y1), (x2,y2), (0,0,255), 3)
img = cv2.line(img, (x1 - (dx * 200), y1 - (dy * 200)),
(x1 + (dx * 200), y1 + (dy * 200)), (0, 255, 0), 2)
cv2.namedWindow('window')
def classical(img):
img = cv2.resize(img, (224, 224)) # For Resnet18
img1 = img.copy()
img = cv2.flip(img, 1)
v_x = int(img.shape[1] / 2)
v_y = int(img.shape[0] / 2)
#Calculate lines using the LSD algorithm
# v_x = int(v_x)
# v_y = int(v_y)
img_gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) # BGR TO GRAYSCALE
img_gray = cv2.medianBlur(
img_gray, 3
) # Removes salt and pepper noise by convolving the image with a (3,3) square kernel
img_gray = cv2.GaussianBlur(
img_gray, (9, 9), 0
) # Smoothens the image by convolving the with a (9,9) Gaussian filter
lines = lsd(img_gray)
# Selecting required lines form all possible lines.
# cv2.line(img,(x3,y3),(x4,y4),(0,0,255),w/2)
sel_lines, thetas, mags, w = process(lines, img, v_x, v_y)
sel_lines = np.array(sel_lines)
print("Number of lines selected are {0}".format(sel_lines.shape[0]))
if sel_lines.shape[0] == 0:
sys.exit("No lines found")
theta_min = sel_lines[:, 5].min()
theta_max = sel_lines[:, 5].max()
mag_max = sel_lines[:, 6].max()
mag_min = sel_lines[:, 6].min()
tmin_ind = np.where(sel_lines == theta_min)[0][0]
tmax_ind = np.where(sel_lines == theta_max)[0][0]
mmin_ind = np.where(sel_lines == mag_min)[0][0]
mmax_ind = np.where(sel_lines == mag_max)[0][0]
L1 = line([sel_lines[tmax_ind, 0], sel_lines[tmax_ind, 1]],
[sel_lines[tmax_ind, 2], sel_lines[tmax_ind, 3]])
L2 = line([sel_lines[tmin_ind, 0], sel_lines[tmin_ind, 1]],
[sel_lines[tmin_ind, 2], sel_lines[tmin_ind, 3]])
R = intersection(L1, L2)
if R:
v_x, v_y = R
else:
sys.exit("No intersection point detected")
innerangle = angle(sel_lines[tmax_ind], sel_lines[tmin_ind])
mvp = np.tan(innerangle)
mvpdeg = (innerangle * 180 / math.pi)
# print 'mvpdeg is {0}'.format(innerangle*180/math.pi)
# print 'Polar line equation value (distance) at the vanishing point is {0}'.format(pm)
cvp = v_y - mvp * (v_x)
yfin = img.shape[0]
xfin = int((yfin - cvp) / mvp)
# cv2.line(img,(sel_lines[tmax_ind,0].astype(int),sel_lines[tmax_ind,1].astype(int)),(sel_lines[tmax_ind,2].astype(int),sel_lines[tmax_ind,3].astype(int)),(255,0,0),int((sel_lines[tmax_ind,4].astype(int))/2))
# cv2.line(img,(sel_lines[tmin_ind,0].astype(int),sel_lines[tmin_ind,1].astype(int)),(sel_lines[tmin_ind,2].astype(int),sel_lines[tmin_ind,3].astype(int)),(0,0,255),int((sel_lines[tmin_ind,4].astype(int))/2))
cv2.line(img, (int(v_x), int(v_y)), (int(xfin), int(yfin)), (0, 255, 0), 4)
cv2.circle(img, (int(v_x), int(v_y)), 10, (0, 255, 255), 3)
# Drawing the desired line in black
cv2.line(img, (int(img.shape[1] / 2), 0),
(int(img.shape[1] / 2), img.shape[0]), (0, 0, 0), 3)
cv2.namedWindow("image")
cv2.setMouseCallback("image", setvp)
# while True:
# dontwrite = 1
# cv2.imshow("image",img)
# cv2.waitKey(1000)
# cv2.destroyAllWindows()
# break
# k = cv2.waitKey(0) & 0xFF
# if k == 27:
# cv2.destroyAllWindows()
# if k == ord('n'):
# break
# if k == ord('s'):
# # output.write(zzz)
# sss = "/home/vdorbala/ICRA/Images/Test/{0}.png".format(file)
# cv2.imwrite(sss,img);
# break
#After setting the new v_x, v_y. The same slope is considered.
# cvp = v_y - mvp*(v_x)
theta_m = thetam(sel_lines[tmax_ind], sel_lines[tmin_ind])
# showagain = 0
# if showagain == 1:
# showagain = 0
# yfin = img.shape[0]
# cv2.line(img,(int(v_x),int(v_y)),(int(xfin),int(yfin)),(0,255,0),int(w/2))
# cv2.circle(img,(int(v_x),int(v_y)), 10, (0,255,255), 3)
# #Drawing the desired line in black
# cv2.line(img,(int(img.shape[1]/2),0),(int(img.shape[1]/2),img.shape[0]),(0,0,0),3)
# cv2.imshow('image',img)
# k = cv2.waitKey(0) & 0xFF
# if k == 27:
# cv2.destroyAllWindows()
# # xfin = int((yfin -cvp)/mvp)
# den = (xfin-v_x)
# if den == 0:
# den = 1
# theta = np.arctan((yfin-v_y)/den)
# # if theta>90:
# if theta<0:
# theta = -math.pi/2 - theta
# if theta>0:
# theta = math.pi/2 - theta
# # innerangle = math.pi/2 - theta
# theta_m = theta
# mvp = np.tan(theta_m)
# # print ("Angle is {}".format(innerangle*180/math.pi))
# cvp = v_y - mvp*(v_x)
# print ("TM is {}".format(theta_m*180/math.pi))
ptd1 = [img.shape[1] / 2, 0]
ptd2 = [img.shape[1] / 2, img.shape[0]]
des_theta = 90 * math.pi / 180 # Because the desired line lies at the center of the image.
des_line = [
img.shape[1] / 2, 0, img.shape[1] / 2, img.shape[0], w / 2, des_theta,
np.sqrt(np.square(ptd2[0] - ptd1[0]) + np.square(ptd2[1] - ptd1[1]))
]
mvp_line = [
v_x, v_y, xfin, yfin, w / 2, innerangle,
np.sqrt(np.square(v_x - xfin) + np.square(v_y - yfin))
]
mdl = np.tan(des_theta)
ydes = ptd1[1]
xdes = ptd1[0]
cdes = ydes - mdl * xdes
v_x = v_x - (img.shape[1] / 2
) ##CHANGING TO CARTESIAN COORDINATES AS GIVEN IN THE PAPER
v_y = -(v_y - (img.shape[0] / 2))
# print (v_x)
v_x = (v_x) * (1 / 5675
) #Converting pixels to meters. Scale is 1m = 5675 pixels
v_y = (v_y) * (1 / 5675)
# print 'theta_m is {0}'.format(theta_m*180/math.pi)
sel_lines = sel_lines.astype(int)
selind = tmax_ind
# print("Thetam is {}".format(theta_m))
h = 1.6 #0.47 for umich
l = 0
w = 0
s = np.sin(theta_m)
c = np.cos(theta_m)
pm = v_x * c + v_y * s
lambda_m = np.cos(theta_m) / h
print("V_X is {}".format(v_x))
error = [[v_x], [theta_m]]
error = np.matrix(error)
# print 'Error is {0}\n'.format(error)
# print (v_x)
# Just testing the ideal case
iderr = ([0], [0])
iderr = np.matrix(iderr)
le = 100 * error
# print ("le is {}".format(le))
lemax = 0
Jw = [[1 + np.square(v_x)],
[((-1) * lambda_m * l * c) + (lambda_m * w * pm) + (pm * s)]]
Jw = np.matrix(Jw)
# print 'Jw is {0}'.format(Jw)
Jv = [[0], [(-1) * (lambda_m * pm)]]
Jv = np.matrix(Jv)
# print '\nJv is {0} \n'.format(Jv)
vconst = 0.2
pinv = (-1) * (np.linalg.pinv(Jw))
# print 'Pseudo inverse is {0} \n'.format(pinv)
fmat = le + Jv * vconst
# print 'fmat is {0},{1},{2} \n'.format(type(fmat),fmat.shape,fmat)
w = pinv * (fmat)
print('w is {0} \n'.format(w))
w = float(w)
return w, img
