Surface Defect Detection of Heat Sink Based on Federated Learning and Improved UNet

• Published in: 2023 China Automation Congress (CAC) pp. 8267 - 8272
• Main Authors: Guo, Feng, Li, Xiaohua, Hou, Zhiyan, Lan, Rukai, Ran, Shaolin, Zhang, Yong
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17-11-2023
• Subjects: Convolution; Data privacy; Data security; Deep Learning; Defect detection; Federated learning; Image segmentation; Quality assessment; Soft sensors; Training
• ISSN: 2688-0938
• DOI: 10.1109/CAC59555.2023.10451277

Surface defect detection directly affects product quality and is essential for industrial production. However, the detection accuracy is limited due to the imbalanced ratio between defective and normal pixels in the samples. In order to address this issue and achieve efficient and accurate heat sink surface defect detection, a federated learning-based Ghost-ECA Separable UNet (FL-GESUNet) model is put forward. Firstly, the GESUNet model is proposed, which contains the Efficient Channel Attention (ECA) module, Ghost module, and separable convolution. This novel combination can reduce the model parameters while maintaining high accuracy. Secondly, in order to overcome the problem of the large gap between the proportion of defective and normal pixels, a hybrid loss function is proposed. The Lovasz Softmax loss and Softmax loss are combined to improve the model's ability to detect defective pixels. Finally, a federated learning (FL) detection framework is introduced to protect data privacy and improve data security. Comparative experiments show that GESUNet achieves an accuracy of 97.99 %, which outperforms the state-of-the-art methods on the heat sink surface defect dataset.