Reproduce “Teacher Encoder-Student Decoder Denoising Guided Segmentation Network for Anomaly Detection” on OpenResearch

About this paper

Shixuan Song, Hao Chen, Shu Hu, Xin Wang, Jinrong Hu, Xi Wu

A self-supervised network, PFADSeg, from Chengdu University of Information Science and Technology, Purdue University, and University at Albany, introduces a teacher-encoder and student-decoder denoising framework for visual anomaly detection and segmentation. It incorporates an enhanced student network with multi-scale feature fusion and a segmentation network with a Rectangular Self-Calibration Module. The model achieved an average image-level AUC of 98.9%, pixel-level AP of 76.4%, and instance-level IAP of 78.7% on the MVTec AD dataset.

Key insights

• Decoupling the teacher and student network architectures, along with training the student on pseudo-anomalous samples in a denoising setup, effectively enables the student to learn more robust and discriminative features for anomalies.
• The Attentional Feature Fusion (AFF) and Parallel Convolutional Attention Recalibration (PCAR) modules significantly enhance the student network's ability to extract and dynamically fuse multi-scale features, improving anomaly detection across different object sizes.
• The Rectangular Self-Calibration Module (RCM) in the segmentation network provides superior localization precision for fine-grained and small anomalies compared to traditional ASPP by creating adjustable rectangular attention regions.
• A combined Focal Loss and L1 Loss effectively addresses class imbalance in pixel-level segmentation while promoting sharper boundaries for anomalous regions.

Summary by alphaXiv.