EAAI Journal 2026 Journal Article
A twin-branch decoupled network for multi-class unsupervised anomaly detection
- Bohan Wang
- Jihong Wan
- Jie Zhao
- Xiaocao Ouyang
- Xiaoping Li
The use of powerful pre-trained Vision Transformer (ViT) encoders in Multi-Class Unsupervised Anomaly Detection (MUAD) can lead to an “identity mapping shortcut”, where the model’s strong generalization inadvertently reconstructs anomalies. To address this specific manifestation of the over-generalization problem, this paper proposes DAD-Net, an innovative hybrid framework combining ViT and Convolutional Neural Networks (CNNs) that imposes synergistic constraints from both the model architecture and the training objective. Architecturally, a novel asymmetric twin-branch CNN decoder is designed to achieve a multi-scale reconstruction of normal patterns. Its shallow branch is specialized for reconstructing high-frequency textures, while its deep branch models abstract semantics. At the objective level, a hard feature loss compels the model to focus on the most complex normal patterns, effectively inhibiting the formation of the “identity mapping shortcut”. Comprehensive experiments validate DAD-Net’s direct applicability to engineering tasks. For industrial defect detection, the framework achieves superior performance on standard benchmarks. Furthermore, the model shows excellent generalization on a challenging cross-domain medical dataset. This highlights its potential as a versatile tool for other critical domains, such as medical diagnostic support. Ablation studies confirm the effectiveness of our core designs, positioning DAD-Net as a robust and practical solution for real-world quality control systems.