Few-Shot Anomaly Classification by Learning to Inpaint Nominal Images

Automated inspection is a crucial aspect in modern industrial manufacturing. Despite its importance, the methods used to perform it suffer from the data scarcity intrinsic to the problem, where only a few anomalous samples are usually available. Recently, fine-tuning of foundation inpainting models has been proposed as a solution. The fine-tuned model is then used to inpaint nominal images where areas corresponding to defective masks have been removed. While effective, random pairing sometimes applies the mask on background or logically unfeasible areas. To counteract this phenomenon, we experiment with generating high-resolution defective images inpainting the few available real defects. Since the resulting images would show limited variance in the non-defective parts, we propose to fine-tune another inpainting model to change nominal parts of generated images. Experimental results on the MVTec-AD dataset demonstrate that our method generates images with complementary properties with respect to those produced by the baseline and training on an ensemble of generated data produces a new state of the art result.