PAEDID: <u>P</u>atch <u>A</u>utoencoder-based <u>D</u>eep <u>I</u>mage <u>D</u>ecomposition for pixel-level defective region segmentation

Unsupervised pixel-level defective region segmentation is an important task in image-based anomaly detection for various industrial applications. The state-of-the-art methods have their own advantages and limitations: matrix-decomposition-based are robust to noise, but lack complex background image modeling capability; representation-based good at localization, accuracy shape contour extraction; reconstruction-based detected match well with the ground truth contour, noisy. To combine best of both worlds, we present unsupervised Patch AutoEncoder-based Deep Image Decomposition (PAEDID) method segmentation. In training stage, learn common as a deep prior by patch autoencoder network. inference formulate decomposition problem sparsity regularizations. By adopting proposed approach, regions can be accurately extracted fashion. We demonstrate effectiveness PAEDID simulation studies dataset case study.