如何防止 IPAINTING 堵塞并改善着色

我想从中删除所有使用 IPAINTING 的文本IMAGE https://i.stack.imgur.com/vxOdS.jpg。我一直在尝试各种方法，最终发现可以通过OCR获得结果，然后使用阈值MASK THE IMAGE。

processedImage = preprocess(partOFimg) 
mask = np.ones(img.shape[:2], dtype="uint8") * 255
for c in cnts:
        cv2.drawContours(mask, [c], -1, 0, -1)
img = cv2.inpaint(img,mask,7,cv2.INPAINT_TELEA)

预处理操作：

  ret,thresh1 = cv2.threshold(gray, 0, 255,cv2.THRESH_OTSU|cv2.THRESH_BINARY_INV)
rect_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 3))
dilation = cv2.dilate(thresh1, rect_kernel, iterations = 1)
edged = cv2.Canny(dilation, 50, 100)
cnts = cv2.findContours(edged.copy(), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
cnts = imutils.grab_contours(cnts)
mask =

np.ones(img.shape[:2], dtype="uint8") * 255

当我运行上面的代码时，我这里是输出图像OUTPUT https://i.stack.imgur.com/C3Q2k.png。正如我们所看到的，它在图像上制作了一些不同颜色的块，我想防止这种情况发生，我该如何实现这一点？我发现很多时候蒙版图像都没有很好地形成，并且在文本为白色的情况下，预处理无法正确进行。 如何防止这些其他颜色的块在图像上形成？

灰色子图像GRAYED https://i.stack.imgur.com/i4AMq.png

阈值子IMG部分：阈值图像 https://i.stack.imgur.com/qLEzE.png

蒙版图像Masked https://i.stack.imgur.com/E22hL.png

EDIT 1:

I've managed to get this new better result by noticing that my threshold is the best mask I can get. After doing this I performed the masking process 3 different times with variable masks and inversions. I did the inpainting algorithm 3 times, it basically the other times inverse the mask, because in some cases required mask is the inversed mask. Still I think it needs improvement, If I chose a different image the results are not so good.

Python/OpenCV 修复方法通常不适合您的图像类型。它们在薄（划痕状）区域而不是大块上效果最好。您确实需要一个示例类型方法，例如https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/criminisi_tip2004.pdf https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/criminisi_tip2004.pdf。但 OpenCV 没有这个。

然而，我怀疑 OpenCV 方法在这里确实有效，因为你填充的是恒定的颜色（绿色）而不是纹理。因此，您最好尝试仅获取字母（字符）的掩码，而不是单词的矩形块。因此，为了向您展示我的意思，这是我的 Python/OpenCV 方法。

Input:

• 读取输入
• 绿色标志上的门槛
• 应用形态学将其关闭并保留为掩模1
• 将蒙版应用于图像以将标志的外部变黑
• 新图像中白色的阈值并保留为 mask2
• 应用形态扩张将其稍微放大并另存为mask3
• 进行修补
• 保存结果

import cv2
import numpy as np

# read input
img = cv2.imread('airport_sign.jpg')

# threshold on green sign
lower = (30,80,0)
upper = (70,120,20)
thresh = cv2.inRange(img, lower, upper)

# apply morphology close
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (135,135))
mask1 = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel)

# apply mask to img
img2 = img.copy()
img2[mask1==0] = (0,0,0)

# threshold on white
#gray = cv2.cvtColor(img2, cv2.COLOR_BGR2GRAY)
#mask2 = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY+cv2.THRESH_OTSU)[1]
lower = (120,120,120)
upper = (255,255,255)
mask2 = cv2.inRange(img2, lower, upper)

# apply morphology dilate
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5,5))
mask3 = cv2.morphologyEx(mask2, cv2.MORPH_DILATE, kernel)


# do inpainting
result1 = cv2.inpaint(img,mask3,11,cv2.INPAINT_TELEA)
result2 = cv2.inpaint(img,mask3,11,cv2.INPAINT_NS)

# save results
cv2.imwrite('airport_sign_mask.png', mask3)
cv2.imwrite('airport_sign_inpainted1.png', result1)
cv2.imwrite('airport_sign_inpainted2.png', result1)

# show results
cv2.imshow('thresh',thresh)
cv2.imshow('mask1',mask1)
cv2.imshow('img2',img2)
cv2.imshow('mask2',mask2)
cv2.imshow('mask3',mask3)
cv2.imshow('result1',result1)
cv2.imshow('result2',result2)
cv2.waitKey(0)
cv2.destroyAllWindows()

Mask 3:

修复 1（Telea）：

修补 2 (NS)：