Shape-guided industrial anomaly detection
Hao Chen, Yaping Yan · IET conference proceedings. · 2025
Current industrial anomaly detection models mainly employ either convolutional neural networks (CNNs) or Transformer networks in isolation, often overlooking their unique strengths and limitations in anomaly detection. This limits the capability of model to effectively handle diverse types of anomaly patterns. To this end, we propose a novel progressive framework for anomaly detection by strategically integrating these two architectures to leverage their respective advantages. First, we utilize the global receptive field and low-frequency information processing capability of the Transformer to restore the shape of input image. This restored shape information is then serves as auxiliary input to a CNN-based reconstructive module, enabling the subnetwork to perform a more refined reconstruction by relying on the structural context. Finally, by comparing the fully reconstructed image to the test image, we can accurately localize anomalous regions. Our approach achieves state-of-the-art performance on industrial anomaly detection datasets, demonstrating its effectiveness in enhancing anomaly detection across a variety of scenarios.