cv：：边缘检测原理，Python代码实现

1. Sobel

Roberts算子：

Prewitt算子：

Sobel算子：

Laplacian算子：

步骤：

1. 灰度图与卷积核相卷

2. 对结果求绝对值

方法2=方法1

#方法1：
sx = cv2.Sobel(img,ddepth=cv2.CV_64F,dx=1,dy=0,ksize=3)
sx = cv2.convertScaleAbs(sx)
#方法2：
kernelx = np.array([[-1,0,1],[-2,0,2],[-1,0,1]],dtype=int)
ix = cv2.filter2D(img,cv2.CV_64F,kernelx)
ix = cv2.convertScaleAbs(ix)
iy = cv2.filter2D(img,cv2.CV_64F,kernely)
iy = cv2.convertScaleAbs(iy)
ixy = cv2.addWeighted(ix, 0.5, iy, 0.5, 0)

2. Canny

步骤：

1. 高斯滤波

2. sobel计算梯度幅值和方向

3. NMS非极大值抑制：8个方向，4条直线，前后直线上最大值则保留，否为0

4. 双阈值检测：>th2保留，<th1去掉，中间看四周有大的就保留

代码就不一一贴了，都放在后面了。

第一步，高斯滤波就是卷积高斯核

第二步，梯度方向和幅值

第三步，NMS，暴力计算，很慢，被去掉了很多

第四步，双阈值检测，很慢

cv2接口运行结果，还是存在一些差别

 imgcanny = cv2.Canny(img,th1,th2,apertureSize=3,L2gradient=True)

代码。

def CannyApi(th1,th2):

    #第一步，高斯模糊
    kernel = cv2.getGaussianKernel(5,1)* cv2.getGaussianKernel(5,1).T
    imgblur = cv2.filter2D(img, cv2.CV_64F, kernel)
    imgblur = cv2.convertScaleAbs(imgblur)

    #第二步，计算梯度幅值和强度，同sobel
    sx = cv2.Sobel(imgblur, ddepth=cv2.CV_64F, dx=1, dy=0, ksize=3)
    sy = cv2.Sobel(imgblur, ddepth=cv2.CV_64F, dx=0, dy=1, ksize=3)

    sxy =np.sqrt(sx*sx+sy*sy)
    sxy = np.array(sxy,dtype=np.uint8)#显示很重要
    sxy[sxy>255] = 255
    cv2.namedWindow("sxy",0)
    cv2.resizeWindow("sxy",640,480)
    cv2.imshow("sxy",sxy)
    # #显示结果不一样dtype=np.uint8，数据对比一下
    # ix = cv2.convertScaleAbs(sx)
    # c = ix-np.sqrt(sx*sx)
    # print(np.max(c),np.min(c))
    #cv2.imshow("ix", ix)
    #方向
    sxyo = np.arctan2(sy,sx)
    cv2.namedWindow("sxyo",0)
    cv2.resizeWindow("sxyo",640,480)
    cv2.imshow("sxyo", sxyo)
   # cv2.waitKey(0)

    ##第三步，NMS非极大值抑制，遍历在它的方向上是局部最大值,四个方向
    sxynms = sxy
    for i in range(1,sxy.shape[0]-1):#rows1167
        for j in range(1,sxy.shape[1]-1):#cols1751
            dir = sxyo[i][j]*180/np.pi#[-90,90]
            dir = dir+180 if dir<0 else dir#[0,180]
            c = sxy[i][j]
            p=255
            n=255
            if dir< 22.5 or dir>157.5:
                p = sxy[i][j+1]
                n = sxy[i][j-1]
            elif dir>=22.5 and dir<67.5:
                p = sxy[i-1][j+1]
                n = sxy[i+1][j-1]
            elif dir >= 67.5 and dir < 112.5:
                p = sxy[i-1][j]
                n = sxy[i+1][j]
            elif dir>=112.5 and dir<=157.5:
                p = sxy[i+1][j+1]
                n = sxy[i-1][j-1]
            if c<p or c<n:
                sxynms[i][j] = 0
    c = sxy - sxynms
    print(np.max(c),np.min(c))
    cv2.namedWindow("sxynms",0)
    cv2.resizeWindow("sxynms",640,480)
    cv2.imshow("sxynms", sxynms)
    #cv2.waitKey(0)
    #第四步，双阈值，> Y,<N,看周围有没有
    sxythresh = sxynms
    for i in range(0, sxynms.shape[0]):  # rows1167
        for j in range(0, sxynms.shape[1]):  # cols1751
            c = sxynms[i][j]
            si  = 0 if i-1<0 else i-1
            ei = sxynms.shape[0]-1 if i+2>sxynms.shape[0]-1 else i+2
            sj = 0 if j-1<0 else j-1
            ej = sxynms.shape[1]-1 if j+2>sxynms.shape[1]-1 else j+2
            s = sxynms[si:ei,sj:ej]
            if np.max(s)>=th2 and c>=th1:
                sxythresh[i][j] = 255
            else:
                sxythresh[i][j] = 0
    cv2.namedWindow("sxythresh",0)
    cv2.resizeWindow("sxythresh",640,480)
    cv2.imshow("sxythresh", sxythresh)

    #API调用
    imgcanny = cv2.Canny(img,th1,th2,apertureSize=3,L2gradient=True)
    c = imgcanny - sxythresh
    print(np.max(c),np.min(c))

    cv2.namedWindow("imgcanny",0)
    cv2.resizeWindow("imgcanny",640,480)
    cv2.imshow("imgcanny", imgcanny)

    cv2.waitKey(0)