DSU-Net: Dynamic Stacked U-Net for Enhancing Mobile Screen Defect Detection

Defect detection in mobile screens is a pivotal aspect of the mobile manufacturing industry. In this paper, we present a novel approach for mobile screen defect detection based on the U-Net network named DSU-Net. We observe that the conventional U-Net encoder is composed of only a few convolutional layers, which impedes the network's capability to comprehensively capture crucial target features. Additionally, the direct connection between the bottom encoder and decoder in the U-shaped structure increases computational complexity, making it unsuitable for real-time requirements. To address these issues, we introduce a dynamic stacking encoder embedded with efficient convolutional modules to replace the conventional U-Net encoder. This dynamic stacking encoder can dynamically adjust the utilization of deep and shallow feature maps based on specific objectives, leading to improved segmentation performance with enhanced accuracy. Furthermore, to simplify the network structure and reduce computational complexity, we introduce a path fusion module. Lastly, we incorporate a boundary loss function into the loss function and adopt a weighted joint training strategy to balance the network's attention between defect boundaries and primary regions. Comparative experiments demonstrate that our proposed framework significantly enhances the network's detection accuracy without introducing excessive parameters.