Inpainting Transformer for Anomaly Detection
This work addresses anomaly detection in computer vision, offering a novel transformer-based approach that improves segmentation performance, though it is incremental as it builds on existing inpainting methods.
The paper tackled anomaly detection in computer vision by framing it as a patch-inpainting problem and proposed the Inpainting Transformer (InTra), a purely self-attention-based method that discards convolutions to integrate information from distant regions. The result showed that InTra, when trained from scratch without extra data, achieved state-of-the-art performance on the MVTec AD dataset for detection and surpassed it for segmentation.
Anomaly detection in computer vision is the task of identifying images which deviate from a set of normal images. A common approach is to train deep convolutional autoencoders to inpaint covered parts of an image and compare the output with the original image. By training on anomaly-free samples only, the model is assumed to not being able to reconstruct anomalous regions properly. For anomaly detection by inpainting we suggest it to be beneficial to incorporate information from potentially distant regions. In particular we pose anomaly detection as a patch-inpainting problem and propose to solve it with a purely self-attention based approach discarding convolutions. The proposed Inpainting Transformer (InTra) is trained to inpaint covered patches in a large sequence of image patches, thereby integrating information across large regions of the input image. When training from scratch, in comparison to other methods not using extra training data, InTra achieves results on par with the current state-of-the-art on the MVTec AD dataset for detection and surpassing them on segmentation.