Improved anomaly detection by training an autoencoder with skip connections on images corrupted with Stain-shaped noise

In industrial vision, the anomaly detection problem can be addressed with an\nautoencoder trained to map an arbitrary image, i.e. with or without any defect,\nto a clean image, i.e. without any defect. In this approach, anomaly detection\nrelies conventionally on the reconstruction residual or, alternatively, on the\nreconstruction uncertainty. To improve the sharpness of the reconstruction, we\nconsider an autoencoder architecture with skip connections. In the common\nscenario where only clean images are available for training, we propose to\ncorrupt them with a synthetic noise model to prevent the convergence of the\nnetwork towards the identity mapping, and introduce an original Stain noise\nmodel for that purpose. We show that this model favors the reconstruction of\nclean images from arbitrary real-world images, regardless of the actual defects\nappearance. In addition to demonstrating the relevance of our approach, our\nvalidation provides the first consistent assessment of reconstruction-based\nmethods, by comparing their performance over the MVTec AD dataset, both for\npixel- and image-wise anomaly detection.\n