Teacher Encoder-Student Decoder Denoising Guided Segmentation Network for Anomaly Detection

TL;DR

This work tackles unsupervised anomaly detection and localization by introducing PFADSeg, a three‑path framework that blends a pre‑trained teacher encoder with an enhanced denoising student decoder and a self‑supervised segmentation network. A key innovation is the PCAR module, which performs parallel multi‑scale feature extraction and attention recalibration to improve regional anomaly localization, combined with an AFF‑based fusion and an RCM for refined context. The training proceeds in two stages using synthetic, pseudo‑anomalous inputs to align student features to the teacher and then fuse their outputs for segmentation, with a loss that combines $L_{focal}$ and $L_{l1}$ to emphasize rare anomalous pixels. On MVTec AD, PFADSeg delivers image‑level AUC of $98.9\%$, pixel‑level AP of $76.4\%$, and competitive instance‑level detection, outperforming several contemporary methods and demonstrating robust segmentation across object and texture categories.

Abstract

Visual anomaly detection is a highly challenging task, often categorized as a one-class classification and segmentation problem. Recent studies have demonstrated that the student-teacher (S-T) framework effectively addresses this challenge. However, most S-T frameworks rely solely on pre-trained teacher networks to guide student networks in learning multi-scale similar features, overlooking the potential of the student networks to enhance learning through multi-scale feature fusion. In this study, we propose a novel model named PFADSeg, which integrates a pre-trained teacher network, a denoising student network with multi-scale feature fusion, and a guided anomaly segmentation network into a unified framework. By adopting a unique teacher-encoder and student-decoder denoising mode, the model improves the student network's ability to learn from teacher network features. Furthermore, an adaptive feature fusion mechanism is introduced to train a self-supervised segmentation network that synthesizes anomaly masks autonomously, significantly increasing detection performance. Rigorous evaluations on the widely-used MVTec AD dataset demonstrate that PFADSeg exhibits excellent performance, achieving an image-level AUC of 98.9%, a pixel-level mean precision of 76.4%, and an instance-level mean precision of 78.7%.

Figures (5)

• Figure 1: The traditional S-T framework (a) and our denoising student-teacher distillation framework (b) in the KD mode.
• Figure 2: Overview of PFADSeg. The training consists of two stages. In stage (a), the improved student network is trained with pseudo-anomalous images (created by blending anomaly masks with normal images), while the teacher network, with fixed weights, processes normal images. The student learns to match features extracted by the teacher. In stage (b), both networks take pseudo-anomalous images as input, with weights fixed. Their normalized outputs are combined via element-wise multiplication and concatenation to form input for the segmentation network, which is trained using the anomaly masks as ground truth.
• Figure 3: Comparison between the original Residual Block (a) in ResNet18 and the proposed PA_Residual Block (b).
• Figure 4: Overall architecture of the PCAR module. The feature map is represented by its dimensions $H$, $W$, and $C$, denoting height, width, and the number of channels, respectively. The Softmax operation recalibrates channel attention weights to capture long-range channel dependencies. The symbol $\otimes$ indicates element-wise multiplication, and channel_sum refers to summation along the channel dimension.
• Figure 5: Visual comparison of anomaly segmentation between our method and DeSTSeg. For each example, first column: input image; second column: Ground Truth; third and fourth columns: predicted segmentation results.