def argsort(self, axis=1, desc=False):
if axis == 0:
sortFlags = cv.SORT_EVERY_COLUMN
elif axis == 1:
sortFlags = cv.SORT_EVERY_ROW
if desc:
sortFlags += cv.SORT_DESCENDING
else:
sortFlags += cv.SORT_ASCENDING
out = Mat(self.shape, dtype=int32, UMat=self.UMat)
cv.sortIdx(self._, sortFlags, out._)
return out
def img_process(self, img_path='', filename='default'):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#cv2.imshow('%s'%img_path, img)
#blur = cv2.GaussianBlur(gray, (5, 5), 0)
ret, bin_img = cv2.threshold(gray, 125, 255, cv2.THRESH_BINARY)
#cv2.imshow('%s bin_img'%img_path, bin_img)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
bin_img = cv2.morphologyEx(bin_img, cv2.MORPH_CLOSE, kernel)
#cv2.imshow("%s close"%img_path, bin_img)
bin_img, contours, hierarchy = cv2.findContours(
bin_img, cv2.RETR_LIST, cv2.CHAIN_APPROX_NONE)
print len(contours)
idx = 0
areas = np.zeros(len(contours))
for con in contours:
areas[idx] = cv2.contourArea(con)
idx += 1
areas_s = cv2.sortIdx(areas,
cv2.SORT_DESCENDING | cv2.SORT_EVERY_COLUMN)
for idx in areas_s:
if areas[idx] < 200:
break
poly_img = np.zeros(bin_img.shape, dtype=np.uint8)
cv2.drawContours(poly_img, contours, idx, [255, 255, 255], -1)
poly_img = poly_img & bin_img
#cv2.imshow('%scontour %d' % (img_path, idx), poly_img)
cv2.imwrite('%s_%d.jpg' % (img_path.split('.')[0], idx), poly_img)
def contours_sort_idx(contours):
"""轮廓排序
"""
# 按面积排序
areas = numpy.zeros(len(contours))
for idx,cont in enumerate(contours):
areas[idx] = cv2.contourArea(cont)
return cv2.sortIdx(areas, cv2.SORT_DESCENDING | cv2.SORT_EVERY_COLUMN)
def nms(self, boxes, iou_thresh):
""" This Function takes the output of yolov3 and loops though all
to know which ones are similar, This function keeps the best
prediction for an object
"""
# If there are no bounding boxes do nothing
if len(boxes) == 0:
return boxes
# Create a PyTorch Tensor to keep track of the detection confidence
# of each predicted bounding box
det_confs = np.zeros(len(boxes))
# Get the detection confidence of each predicted bounding box
for i in range(len(boxes)):
det_confs[i] = boxes[i][4]
# Sort the indices of the bounding boxes by detection confidence value in descending order.
# We ignore the first returned element since we are only interested in the sorted indices
# print(f'det_confs --> {det_confs}')
# _,sortIds = torch.sort(det_confs, descending = True)
sortIds = cv2.sortIdx(det_confs, -1).reshape(-1)
# print(f'sort --> {sortIds}')
# print(f'_ and sortid {_} and {sortIds}')
# Create an empty list to hold the best bounding boxes after
# Non-Maximal Suppression (NMS) is performed
best_boxes = []
# Perform Non-Maximal Suppression
for i in range(len(boxes)):
# Get the bounding box with the highest detection confidence first
box_i = boxes[sortIds[i]]
# Check that the detection confidence is not zero
if box_i[4] > 0:
# Save the bounding box
best_boxes.append(box_i)
# Go through the rest of the bounding boxes in the list and calculate their IOU with
# respect to the previous selected box_i.
for j in range(i + 1, len(boxes)):
box_j = boxes[sortIds[j]]
# If the IOU of box_i and box_j is higher than the given IOU threshold set
# box_j's detection confidence to zero.
if self.boxes_iou(box_i, box_j) > iou_thresh:
box_j[4] = 0
return best_boxes
def pixelOfDarkChannelImage(img, ratio): h, w = img.shape[: 2] numPixels = h * w numSelectPixels = int(numPixels * ratio) newImg = np.zeros(h*w) pixels = [] k = 0 newImg = img.reshape(1, h*w)[0] pixelIndexs = cv2.sortIdx(newImg, cv2.SORT_EVERY_COLUMN + cv2.SORT_DESCENDING) for i in range(numSelectPixels): rowOfPixel = pixelIndexs[i] / w colOfPixel = pixelIndexs[i] % w pixels.append([rowOfPixel, colOfPixel]) return pixels, numSelectPixels
def saltErase(img, contours):
# 按面积排序
areas = np.zeros(len(contours))
idx = 0
for contour in contours:
areas[idx] = cv2.contourArea(contour)
idx = idx + 1
areas_s = cv2.sortIdx(
areas, cv2.SORT_DESCENDING | cv2.SORT_EVERY_COLUMN)
imgClear = np.zeros(img.shape, dtype=np.uint8)
# for idx in areas_s:
# if areas[idx] < 800:
# break
# # 绘制轮廓图像，通过将thickness设置为-1可以填充整个区域，否则只绘制边缘
# cv2.drawContours(imgClear, contours, idx, [255, 255, 255], -1)
cv2.drawContours(imgClear, contours, areas_s[0],
[255, 255, 255], -1)
imgClear = imgClear & img
return imgClear
print("rects[%i] = [(%3d,%3d) from (%3d,%3d)] %5d" %
(i, x, y, w, h, a))
print()
rands = np.zeros((5, 5), np.uint16) # 5행 4열 행렬 생성
starts = cv2.randn(rands[:, :2], 100, 50) # 0~4행까지 시작좌표 랜덤 생성
ends = cv2.randn(rands[:, 2:-1], 300, 50) # 5~9행까지 종료좌표 랜덤 생성
sizes = cv2.absdiff(starts, ends) # 시작좌표와 종료좌표간 차분 절대값
areas = sizes[:, 0] * sizes[:, 1]
rects = rands.copy()
rects[:, 2:-1] = sizes
rects[:, -1] = areas
idx = cv2.sortIdx(areas, cv2.SORT_EVERY_COLUMN).flatten()
# idx = np.argsort(areas, axis=0)
print_rects(rects)
print_rects(rects[idx.astype('int')])
## 리스트 생성 방식
# rects = ["[(%3d,%3d) from (%3d,%3d)]" %(p[0], p[1], s[0], s[1])
# for p, s in zip(starts, sizes)] # 시작좌표와 크기로 리스트 생성
# areas = [s[0]*s[1] for s in sizes] # 넓이 계산 및 리스트에 저장
#
# # 정렬 후, 정렬 원소의 원본 좌표 반환
# sort_idx = cv2.sortIdx(np.array(areas), cv2.SORT_EVERY_COLUMN)
# sort_idx = map(int , sort_idx)
#
# print("-" * 46) # 라인 출력
from PIL import Image
from pylab import *
import cv2
img = cv2.imread('018.jpg')
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
ret, binary = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY)
contoures, hierarchy = cv2.findContours(binary, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_NONE)
areas = np.zeros(len(contoures))
idx = 0
for cont in contoures:
areas[idx] = cv2.contourArea(cont)
idx = idx + 1
areas_s = cv2.sortIdx(areas, cv2.SORT_DESCENDING | cv2.SORT_EVERY_COLUMN)
print areas_s
# (b8, g8, r8) = cv2.split(img)
#
# # 对每个区域进行处理
# for idx in areas_s :
# if areas[idx] < 100 :
# break
#
# # 绘制区域图像，通过将thickness设置为-1可以填充整个区域，否则只绘制边缘
# poly_img = np.zeros(img.shape, dtype = np.uint8 )
# cv2.drawContours(poly_img, contoures, idx, [255,255,255], -1)
# poly_img = poly_img & img
#
# # 得到彩色的图像
def save_png(self, src, imgray, contours, home_path):
width = src.shape[1]
height = src.shape[0]
# 创建存储轮廓点的文件夹
if not os.path.exists(home_path + "\\contours"):
os.mkdir(home_path + "\\contours")
# 创建存储图片的文件夹
if not os.path.exists(home_path + "\\img"):
os.mkdir(home_path + "\\img")
# 创建存储相对坐标的文件夹
if not os.path.exists(home_path + "\\offset"):
os.mkdir(home_path + "\\offset")
# 按面积排序
areas = np.zeros([len(contours)])
index = 0 # 面积的下标指针
# 遍历轮廓，计算每个轮廓的面积
for contour in contours:
areas[index] = cv2.contourArea(contour)
index = index + 1
# 对轮廓按面积排序，sortIdx能够得到这些值在原数组中的序号
areas_s = cv2.sortIdx(areas,
cv2.SORT_DESCENDING + cv2.SORT_EVERY_COLUMN)
(b8, g8, r8, a8) = cv2.split(src)
# 对每个区域进行处理
for index in areas_s:
if areas[index] < width * height * 0.001 or areas[
index] > width * height * 0.1:
continue
m = cv2.moments(contours[index.tolist()[0]])
# 计算中心矩
cx = int(m['m10'] / m['m00'])
cy = int(m['m01'] / m['m00'])
offset_x = (m['m10'] / m['m00']) / height
offset_y = (m['m01'] / m['m00']) / width
# 获得轮廓的边缘点
np.savetxt('{0}\\contours\\{1},{2}'.format(home_path, str(cx),
str(cy)),
(contours[index.tolist()[0]]).reshape((-1, 2)),
fmt='%d',
delimiter=',')
# 获得相对坐标
contours_points = (contours[index.tolist()[0]]).reshape((-1, 2))
x_ = contours_points[:, 0]
y_ = contours_points[:, 1]
x_ = x_ / height
y_ = y_ / width
x_ = x_[:, np.newaxis]
y_ = y_[:, np.newaxis]
contours_points = np.concatenate((x_, y_), axis=1)
np.savetxt(home_path + '\\offset\\' + str(cx) + ',' + str(cy),
contours_points,
fmt='%f',
delimiter=',')
# 绘制区域图像，通过将thickness设置为-1可以填充整个区域，否则只绘制边缘
temp_img = np.zeros(imgray.shape, dtype=np.uint8)
cv2.drawContours(temp_img, contours, index, [255, 255, 255], -1)
# 结合灰度图掩膜
# temp_img = temp_img & imgray
# 得到彩色的图像
color_img = cv2.merge(
[b8 & temp_img, g8 & temp_img, r8 & temp_img, a8 & temp_img])
# cv2.imshow("img",color_img) 为什么会闪退？
cv2.imwrite(
home_path + '\\img\\' + str(cx) + ',' + str(cy) + '.png',
color_img)
# with cv2 # sort : row , ascending sort1 = cv2.sort(matrix, cv2.SORT_EVERY_ROW) # sort : row , descending sort2 = cv2.sort(matrix, cv2.SORT_EVERY_ROW + cv2.SORT_DESCENDING) # sort : col , ascending sort3 = cv2.sort(matrix, cv2.SORT_EVERY_COLUMN) # sort : col , descending sort4 = cv2.sort(matrix, cv2.SORT_EVERY_COLUMN + cv2.SORT_DESCENDING) # with Numpy # sort : x axis sort5 = np.sort(matrix, axis=1) # sort : y axis sort6 = np.sort(matrix, axis=0) # return sort Index - 행렬의 원소를 직접 정렬하지 않고, 정렬된 원소의 원 Index를 반환한다. # cv2 Idx_sort1 = cv2.sortIdx(matrix, cv2.SORT_EVERY_ROW) ''' [[2 0 1 4 3] [4 1 2 3 0] [1 0 3 4 2]] ''' Idx_sort2 = cv2.sortIdx(matrix, cv2.SORT_EVERY_COLUMN) # numpy Idx_sort3 = np.argsort(matrix, axis=0) # y axis
import numpy as np, cv2
m = np.random.randint(0, 100, 15).reshape(3, 5) # 임의 난수 생성
m_sort1 = cv2.sortIdx(m, cv2.SORT_EVERY_ROW) # 행렬 원소의 원본 좌표
m_sort2 = cv2.sortIdx(m, cv2.SORT_EVERY_COLUMN)
m_sort3 = np.argsort(m, axis=0) # 세로축 정렬
print("[m1] = \n%s\n" % m)
print("[m_sort1] = \n%s\n" % m_sort1)
print("[m_sort2] = \n%s\n" % m_sort2)
print("[m_sort3] = \n%s\n" % m_sort3)