Industrial Anomaly Detection using Vision Based Cross Transformers

Few-shot anomaly detection (FSAD) using registration-based techniques for the identification of aberration in image and video feeds has attracted a great lot of researchers in the field. Introduction of transformer with the siamese network has made metric-based learning promising in the identification of anomalies in image streams with only a few shots of comparative support sets. Inspired by how humans use their comparative analysis for identification of differences in images, a registration-based anomaly detection system is introduced leveraging "Fully Cross Transformer" as the backbone architecture with a Siamese network encoder for feature-level registration and finally, loss estimation using Gaussian distribution for achieving probabilistic representation of the normal class. The proposed method incorporates a cross-attention The mechanism for a two-branch-based FSAD model with different batch sizes. Multilevel interactions for feature aggregations are used in the model for obtaining low-level feature representations. Extensive experimentation on various datasets have been conducted with promising results including MVTec AD, Retinal OCT, and Brain MRI. The application for the model includes the identification of Industrial Manufacturing, Disease identification, Agriculture, Earth Resource Management, and other fields.