An Improved Neural Network Based on UNet for Surface Defect Segmentation

Surface defect defection is a critical task in product quality assurance in industrial production lines. The defect characteristics of surface are easily affected by light. A surface defect detection model based on improved UNet-based neural network, surface defect UNet (S-UNet), is proposed to address the limitations of traditional detection algorithms in surface defect detection, as well as the problems of low accuracy, low precision, and cumbersome detection process. S-UNet uses atrous spatial pyramid pooling (ASPP) to acquire different receptive fields, overcoming the limitations of the UNet model with a single receptive field and improving the ability to segment targets of different sizes. S-UNet utilizes features at different levels to achieve complementary information and improve the coherence and accuracy of segmented areas. The model performs well on small industrial commutator datasets and steel surface defect dataset, reaching the optimal level in terms of precision, recall, and F-score. Compared to UNet and other models, as well as traditional methods, the proposed method achieves better results.