Based on your image and your code.This is some advice for you：
lighting condition and image preprocessing: try Gamma correction to improve object lightness.
Binary: do not using Sobel and try to cv2.threshold with Ostu.
Blob analyze：you can filter the real defects by Blob's area,arclength,height,width and so on.It's easy for your solution that using cv2.boundingRect to get the roi of blobs,then calculate the mean of pixel gray of region.
Those printing defects which shown on your image,the average pixel value of the area where they are located will be significantly different from the background.