Hough圆变换的原理很多博客都已经说得非常清楚了,但是手动实现的比较少,所以本文直接贴上手动实现的代码。
这里使用的图片是一堆硬币:
?首先利用通过计算梯度来寻找边缘,代码如下:
def detect_edges(image):
h = image.shape[0]
w = image.shape[1]
sobeling = np.zeros((h, w), np.float64)
sobelx = [[-3, 0, 3],
[-10, 0, 10],
[-3, 0, 3]]
sobelx = np.array(sobelx)
sobely = [[-3, -10, -3],
[0, 0, 0],
[3, 10, 3]]
sobely = np.array(sobely)
gx = 0
gy = 0
testi = 0
for i in range(1, h - 1):
for j in range(1, w - 1):
edgex = 0
edgey = 0
for k in range(-1, 2):
for l in range(-1, 2):
edgex += image[k + i, l + j] * sobelx[1 + k, 1 + l]
edgey += image[k + i, l + j] * sobely[1 + k, 1 + l]
gx = abs(edgex)
gy = abs(edgey)
sobeling[i, j] = gx + gy
# if you want to imshow ,run codes below first
# if sobeling[i,j]>255:
# sobeling[i, j]=255
# sobeling[i, j] = sobeling[i,j]/255
return sobeling
需要注意的是,这里使用的kernel内的数值比较大,所以得到了结果图中的某些位置的数值超过255,但并不影响显示,但如果想通过cv2.imshow来显示,就需要将超过255的地方设为255即可(已经在代码中用注释标出),结果如下:
?接下来就是要进行Hough圆变换,先看代码:
def hough_circles(edge_image, edge_thresh, radius_values):
h = edge_image.shape[0]
w = edge_image.shape[1]
# print(h,w)
edgimg = np.zeros((h, w), np.int64)
for i in range(h):
for j in range(w):
if edge_image[i][j] > edge_thresh:
edgimg[i][j] = 255
else:
edgimg[i][j] = 0
accum_array = np.zeros((len(radius_values), h, w))
# return edgimg , []
for i in range(h):
print('Hough Transform进度:', i, '/', h)
for j in range(w):
if edgimg[i][j] != 0:
for r in range(len(radius_values)):
rr = radius_values[r]
hdown = max(0, i - rr)
for a in range(hdown, i):
b = round(j+math.sqrt(rr*rr - (a - i) * (a - i)))
if b>=0 and b<=w-1:
accum_array[r][a][b] += 1
if 2 * i - a >= 0 and 2 * i - a <= h - 1:
accum_array[r][2 * i - a][b] += 1
if 2 * j - b >= 0 and 2 * j - b <= w - 1:
accum_array[r][a][2 * j - b] += 1
if 2 * i - a >= 0 and 2 * i - a <= h - 1 and 2 * j - b >= 0 and 2 * j - b <= w - 1:
accum_array[r][2 * i - a][2 * j - b] += 1
return edgimg, accum_array
其中输入是我们之前得到的边缘图,以及确定强边缘的阈值,以及一个包含着我们估计的半径的数组;返回值是强边缘图以及参数域矩阵。代码中首先遍历边缘图,通过阈值留下那些较强的位置,这里的阈值需要自己根据自己的输入图进行调节。接着就是进行Hough变换,这里的候选半径集合需要根据自己的输入图进行调节。在绘制参数域的过程中,只遍历了所需正方形区域(大小为?r*r)的?1/4,这是因为在坐出参数域上的一个点之后,由于圆的对称性,就可以找到与之对称的另外三个点,无需额外进行遍历。
最后一步就是从参数域矩阵中提取出结果圆,代码如下,其中筛选阈值需要根据你的输入图像自己调节:
def find_circles(image, accum_array, radius_values, hough_thresh):
returnlist = []
hlist = []
wlist = []
rlist = []
returnimg = deepcopy(image)
for r in range(accum_array.shape[0]):
print('Find Circles 进度:', r, '/', accum_array.shape[0])
for h in range(accum_array.shape[1]):
for w in range(accum_array.shape[2]):
if accum_array[r][h][w] > hough_thresh:
tmp = 0
for i in range(len(hlist)):
if abs(w-wlist[i])<10 and abs(h-hlist[i])<10:
tmp = 1
break
if tmp == 0:
#print(accum_array[r][h][w])
rr = radius_values[r]
flag = '(h,w,r)is:(' + str(h) + ',' + str(w) + ',' + str(rr) + ')'
returnlist.append(flag)
hlist.append(h)
wlist.append(w)
rlist.append(rr)
print('圆的数量:', len(hlist))
for i in range(len(hlist)):
center = (wlist[i], hlist[i])
rr = rlist[i]
color = (0, 255, 0)
thickness = 2
cv2.circle(returnimg, center, rr, color, thickness)
return returnlist, returnimg
注意一下在这一步中需要将那些圆心相近的圆剔除掉,只保留一个结果。
接着是main函数,这没啥好说的:
def main(argv):
img_name = argv[0]
img = cv2.imread('data/' + img_name + '.png', cv2.IMREAD_COLOR)
# print(img.shape[0], img.shape[1])
gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# print(gray_image.shape[0], gray_image.shape[1])
img1 = detect_edges(gray_image)
cv2.imwrite('output/' + img_name + "_after_find_detect.png", img1)
thresh = 1500
# 需要注意的是,在img1中有些地方的像素值是高于255的,这是由于之前的kernel内的数更大
# 但这并不影响图像的显示
# 因此这里的thresh要大于255
radius_values = []
for i in range(10):
radius_values.append(20 + i)
edgeimg, accum_array = hough_circles(img1, thresh, radius_values)
cv2.imwrite('output/' + img_name + "_after_binary.png", edgeimg)
# Findcircle
hough_thresh = 70
resultlist, resultimg = find_circles(img, accum_array, radius_values, hough_thresh)
print(resultlist)
cv2.imwrite('output/' + img_name + "_circles.png", resultimg)
if __name__ == '__main__':
sys.argv.append("coins")
main(sys.argv[1:])
# TODO
下面是我的运行结果:
?
|