Diffusion-Based Image Generation for In-Distribution Data Augmentation in Surface Defect Detection.

In this study, we show that diffusion models can be used in industrialscenarios to improve the data augmentation procedure in the context of surfacedefect detection. In general, defect detection classifiers are trained onground-truth data formed by normal samples (negative data) and samples withdefects (positive data), where the latter are consistently fewer than normalsamples. For these reasons, state-of-the-art data augmentation procedures addsynthetic defect data by superimposing artifacts to normal samples. This leadsto out-of-distribution augmented data so that the classification system learnswhat is not a normal sample but does not know what a defect really is. We showthat diffusion models overcome this situation, providing more realisticin-distribution defects so that the model can learn the defect's genuineappearance. We propose a novel approach for data augmentation that mixesout-of-distribution with in-distribution samples, which we call In Out. Theapproach can deal with two data augmentation setups: i) when no defects areavailable (zero-shot data augmentation) and ii) when defects are available,which can be in a small number (few-shot) or a large one (full-shot). We focusthe experimental part on the most challenging benchmark in thestate-of-the-art, i.e., the Kolektor Surface-Defect Dataset 2, defining the newstate-of-the-art classification AP score under weak supervision of .782. Thecode is available at https://github.com/intelligolabs/in_and_out.