Privacy-preserving small target defect detection of heat sink based on DeceFL and DSUNet

Abstract

Surface defect detection is directly related to the quality of products and crucial for industrial production. Currently, convolutional neural network (CNN)-based structures have been widely used for surface defect detection. However, the locality restriction of convolutional operation poses challenges for surface defect detection, particularly when it comes to detecting small target defects. In addition, there is a risk of leakage during data communication arising from participant attacks. To address the above issues, the Dilated Swin Transformer UNet (DSUNet) model with privacy protection is proposed in this paper. Firstly, the DSUNet model adapting a hybrid CNN-Transformer architecture is designed to effectively address the challenge of detecting small defects, which can capture global and remote semantic information. Secondly, a decentralized federated learning framework (DeceFL) is introduced to protect data privacy. Finally, in order to enhance the interpretability of the model, the regional focus of the defect detection network is visualized through the Grad-CAM method. Comprehensive experiments on the heat sink surface defect dataset are conducted to demonstrate the effectiveness of our proposed model in the field of surface defect detection. The DSUNet achieves an accuracy of 97.98% on the dataset of heat sink surface defect, outperforming the state-of-the-art methods.