如何更有效地预处理以下图像？

Question

I am trying to preprocess images of cloth (ex. below

normal image )

and apply to preprocess it so it would look like this so stitches and rest of the fabric are easily distinguishable (ex. below)

binarised image

My end goal is to train a model on these preprocessed images to detect bad stitches.

The ideal preprocessing should be

step 1:

initial image

step 2:

grayscale

step 3:

noise reduction

step 4:

final binarised image

Currently, I am using a hideous mish-mash of OpenCV functions to achieve the result I posted above; which is bad and not good enough as I want to create a live system. I would like to hear your insights on how to optimise code below

I don't have enough reputation to post images that's why I've posted hyperlinks:)

image = cv2.imread("C:/Users/name/Downloads/timepass/1.jpg")
image = cv2.resize(image, (800, 800))
thresh = 185
blur = cv2.bilateralFilter(image, 50, 175, 175)
blur = cv2.bilateralFilter(blur, 50, 175, 175)
im_bw = cv2.threshold(blur, thresh, 255, cv2.THRESH_BINARY)[1]
gray = cv2.cvtColor(im_bw, cv2.COLOR_BGR2GRAY)
final = cv2.threshold(gray, 50, 255, cv2.THRESH_BINARY)[1]
cv2.imshow("bi", final)

Answer 1

I recommend that the first processing you apply is conversion to grayscale. This will shorten the processing time, so that the following functions will only process one channel instead of three.

For noise removal, the bilateral filter is computationally expensive, its advantage is that it smoothes the image without losing edge information. However, I would recommend a Gaussian filter, or averaging, with small kernels to avoid losing too much information.

In binarization, you can use an OTSU or adaptive threshold, they obtain dynamic thresholds according to the image. Try them and see how it works. Then you apply a median filter to remove the impulsive noise, present in the binarized image. After that you use morphological operations, I recommend you to use aperture especially useful for noise removal. You can also play with dilation to increase the size of the stitches.

I made a small program, you can include the morphological operations and add the otsu threshold. The function "remove_blobs_on_edges" as the name suggests removes those blobs that are on the edges. The program is as follows:

import numpy as np
import cv2 
 
def remove_blobs_on_edges(image):
   
    """
        Eliminates blobs on edges

    """

    # use the binarized image as a mask for flood filling
    activation = np.array(image, dtype=np.uint8)
    
   
    mask_shape = (image.shape[0] + 2, image.shape[1] + 2)
    ffill_mask = np.zeros(mask_shape, dtype=np.uint8)
    # top and bottom rows
    for i in range(0, activation.shape[0]):
        for j in [0, activation.shape[1] - 1]:
            if activation[i,j]:
                cv2.floodFill(activation, ffill_mask, (j, i), 0)

    # left and right columns
    for i in [0, activation.shape[0] - 1]:
        for j in range(0, activation.shape[1]):
            if activation[i,j]:
                cv2.floodFill(activation, ffill_mask, (j, i), 0)
    

    return activation


image = cv2.imread('image2.png',0)
blured = cv2.blur(image,(3,3))
thresh = cv2.threshold(blured, 150, 255, cv2.THRESH_BINARY)[1]
thresh = remove_blobs_on_edges(thresh)

blured_mask = cv2.medianBlur(thresh,3)
before_after = np.hstack((thresh, blured_mask))

cv2.imshow("Result- Before/After",before_after)
cv2.waitKey(0)
cv2.destroyAllWindows()

Result: