DMU-TransNet: Dense multi-scale U-shape transformer network for anomaly detection

Anomaly detection is crucial in medical and industrial sectors. Despite the promising results of Convolutional Neural Networks (CNNs), existing approaches often focus predominantly on local visual features, neglecting global and multi-scale features. This limitation leads to unstable training and suboptimal detection performance. To address these challenges, this paper introduces a novel anomaly detection model, named DMU-TransNet. Firstly, the Convolution with Transformer (ConvTR) module is designed to enlarge the receptive field and enhance the model's global dependence. Secondly, the Dense Multi-scale Skip Connection (DMSC) module amalgamates multi-scale features from various layers, enabling the model to comprehend local details and global context simultaneously, while mitigating the vanishing gradients problem. Finally, the Multi-level Fusion (MF) module merges multi-scale features from multi-level decoders, minimizing information loss and enabling effective capture of diverse information. Extensive experimental results confirm that our method achieves optimal performance on several mainstream datasets. Our code is available at https://github.com/ml-AD/DMU-TransNet.