DualFormer: dual-branch transformer for industrial defect detection

Industrial surface defect detection technology is a crucial aspect of industrial automation. However, existing methods often struggle to balance the detection of structural errors in images with the ability to distinguish subtle defects. To address this issue, this paper proposes a novel dual-branch Transformer defect detection algorithm called DualFormer. This framework consists of two independent branches: one branch focuses on the segmentation of RGB images, emphasizing the ability to distinguish subtle anomalous regions, while the other utilizes high-frequency domain information for image restoration, concentrating on the detection of structural defects in images. Both branches employ the same Transformer encoder to extract feature structures but utilize different decoder designs, each optimized for segmentation and restoration tasks, respectively. The weights of the feature extraction network are not shared during the training process, ensuring that each branch can fully learn task-specific feature representations. Finally, the two branches are fused through weighted averaging, enabling the model to precisely detect defects in anomalous images, thereby enhancing the overall defect detection capability. Comparative experiments on the MVTec AD benchmark dataset show that this method achieves a 0.4% improvement in pixel-level AUROC over the baseline method FAIR.