Dual-channel image anomaly detection
Yongxin Jiang, Yaping Yan · IET conference proceedings. · 2025
Image anomaly detection focuses on identifying areas that deviate from normal patterns. The most commonly used networks usually follow an encoder-decoder architecture. These networks learn from normal training samples to predict masked information and use the reconstruction error related to the masked data as an abnormality score. However, previous approaches ignore that the size of surrounding areas required to reconstruct various anomalies are different. Small receptive fields lack a holistic understanding of shape and spatial relationships, while large receptive fields tend to involve more undesired information and be ill-suited for handling local structures and textures. Based on these insights, this paper proposes an edge guided anomaly detection method. Specifically, we first utilize an encoder-decoder network with residual block and large receptive field to achieve edge restoration. Then a shallow network with small receptive field is introduced to conduct edge guided image reconstruction. Finally, we learn a discriminative network to identify anomalies, leveraging both of structural and appearance discrepancies for more comprehensive and accurate detection. Experiments on the industrial inspection benchmark dataset demonstrate that our method achieves state-of-the-art performance.