sobel边缘检测

sobel边缘检测 Sobelfilter numpy

def Sobelfilter(m):
    kernelx = [[1, 2, 1], [0, 0, 0], [-1, -2, -1]]
    kernely = [[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]
    # m2=scipy.ndimage.filters.convolve(m, kernelx, mode='nearest')
    m2 = ndimage.filters.convolve(m, kernelx, mode='nearest')
    m3 = ndimage.filters.convolve(m, kernely, mode='nearest')#numpy
    #print('m2',m2,m3)
    #需要加个abs
    #return  torch.from_numpy(m2+m3)
    return  (np.abs(m2+m3))

使用F.conv2d tensor

def sobel(batch_img):
    #print(batch_img.shape)#torch.Size([1, 1, 5, 5])
    ksize=3
    C = 1
    # kernelx = torch.tensor([[1.0, 2, 1], [0, 0, 0], [-1, -2, -1]])#.cuda()
    # kernely =  torch.tensor([[-1.0, 0, 1], [-2, 0, 2], [-1, 0, 1]])#.cuda()

    kernelx = torch.tensor([[-1.0, -2, -1], [0, 0, 0], [ 1,  2,  1]])  # .cuda()
    kernely = torch.tensor([[1.0, 0, -1], [2, 0, -2], [1, 0, -1]])  # .cuda()
    kernely = kernely.view(1, 1, ksize, ksize).repeat(C, 1, 1, 1)#.cuda()
    kernelx = kernelx.view(1, 1, ksize, ksize).repeat(C, 1, 1, 1)#.cuda()
    #print(kernely.shape)
    pad = (ksize - 1) // 2  # 保持卷积前后图像尺寸不变

    #batch_img_pad=batch_img
    #batch_img_pad = F.pad(batch_img, pad=[pad, pad, pad, pad], mode='reflect')
    batch_img_pad = F.pad(batch_img, pad=[pad, pad, pad, pad], mode='replicate')#’constant‘, ‘reflect’ or ‘replicate’
    #print('batch_img_pad',batch_img_pad)
    weighted_pix = F.conv2d(batch_img_pad, weight=kernelx, bias=None,
                            stride=1, padding=0, groups=C)
    weighted_piy = F.conv2d(batch_img_pad, weight=kernely, bias=None,
                            stride=1, padding=0, groups=C)
    #print('weighted_pix',weighted_pix,weighted_piy)
    #return weighted_pix+weighted_piy


    return torch.abs(weighted_pix + weighted_piy)

test

def testsobel():
    emeloss = EME()
    lambdaR = 0.299
    lambdaG = 0.587
    lambdaB = 0.114
    path=r'F:traindata_copyLOLeval_normal/normal00690.png'
    img=cv2.imread(path,cv2.IMREAD_GRAYSCALE)
    im=Sobelfilter(img)
    img = np.expand_dims(img, axis=2)
    source2 =torch.from_numpy(np.ascontiguousarray(img)).permute(2, 0, 1).float()#.div(255.)# util.uint2tensor3(img)
    source2 = torch.unsqueeze(source2, dim=0)
    img = source2 #* 255.0
    print(img.shape)
    #img=img.cuda()
    im2=sobel(img)
    print(img.shape,im2.shape)
    im2 = im2.data.squeeze().float().cpu().numpy()
    cv2.imwrite(r'F:traindata_copyLOL/normal00690.png',im)
    cv2.imwrite(r'F:traindata_copyLOL/normal006902.png',im2)
    r, g, b = readgray(path )
    r=np.array(r)
    g=np.array(g)
    b=np.array(b)
    cv2.imwrite(r'F:traindata_copyLOL/normal00690r.png', r)
    cv2.imwrite(r'F:traindata_copyLOL/normal00690g.png', g)
    cv2.imwrite(r'F:traindata_copyLOL/normal00690b.png', b)
    lo1 = lambdaR * (r) + lambdaG * (g) + lambdaB * (b)
    cv2.imwrite(r'F:traindata_copyLOL/normal00690rgb.png', lo1)
    print(lo1.shape)

输入

Sobelfilter

R通道

G通道

B通道

RGB加权

sobel 灰度图