The images are either 32×32 pixels or 64×64 pixels in size. There are different types of classes: ‘good’, ‘color’, ‘cut’, ‘hole’, ‘thread’, and ‘metal contamination’. Each image can be rotated in eight different ways: 0, 20, 40, 60, 80, 100, 120, or 140 degrees. There are separate sets of images for training and testing. The patches in these datasets are randomly generated from source images, and there’s no overlap between the source images used for training and testing.
different tasks are possible:
Classify the types of defects.
Predict the angle of rotation, but only using images of “good” quality for training and testing on other types of defects.
Learn about the texture or patterns in the images without needing labels, a bit like teaching yourself.