Dlib C++ Metal part defect Training

I am trying to find the best way to train images where the entire image is covert by the object. The object is a metal part with a small patern. I would like to detect scratches or other part defects.

I already did some tests with MMOD + Part detection (using 25 Points)MMOD + Part detection.

I also did some tastings with SVM so that i could avoid to use a GPU and still have a nice detection time.

In both cases I obtain the score of the results, and the detection works quite well. (even if its a small data set).

My doubt is, if i chose the window to occupy the whole height of the image during the training (in SVM and DNN), and if i increase the dataset so that i use 1 or 2 k of images, could i detect scratches, or other defects?

Another question is, if I detect these defects, could i visualize them.