Dual-Branch Knowledge Distillation via Residual Features Aggregation Module for Anomaly Segmentation

Unsupervised image anomaly detection and segmentation algorithms are of great significance in the actual industrial quality inspection process. The anomaly detection method, based on the knowledge distillation framework, detects anomalies by characterizing the differences between abnormal samples. This method uses similar model structures to construct the teacher and student networks, which leads to limited model representation ability and the problem of detection failure due to the disappearance of abnormal activation values. This article solves the problem of disappearing abnormal responses by improving the diversity of model representation. The multiscale input reconstruction branch in the dual-branch knowledge distillation (DBKD) model proposed in this article improves its representation ability by restoring the representation of the input at multiple scales. The multiscale feature information extraction branch enhances its ability to capture defect detail information by extracting feature information at different scales. In addition, we design a residual feature aggregation module (RFAM) to condense the high-dimensional features of the teacher model into compact and effective low-dimensional feature embeddings, which ensures the effectiveness of the multiscale input reconstruction branch input. The proposed DBKD achieves the latest state-of-the-art (SOTA) on the famous MVTec AD dataset, with an area under the receiver operating characteristic curve (ROCAUC) of 98.1% for anomaly detection and 98.2% for anomaly segmentation in all 15 categories. Our code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/EWAN9709/DBKD</uri>.