An improved detection algorithm using RFS-YOLOv8n for wafer surface defect

Abstract Wafer defect detection is crucial for timely identification of broken dies and ensuring chip quality. However, due to the presence of numerous intricate circuit patterns in wafer images and limited platform resources, the majority of existing wafer inspection models exhibit low accuracy, making it challenging to effectively balance detection performance and resource consumption. To address this issue, an enhanced defect detection algorithm for wafer surface, named RFS-YOLOv8n, was proposed in this paper. Firstly, the C2f structure of the YOLOv8n model was optimized by integrating RepViTBlock and Context anchor attention, which reduces feature loss and improves feature extraction capability. Additionally, a new Feature-Focused Diffusion Pyramid Network was introduced, which provides detailed contextual information at each scale through unique feature-focused and diffusion modules. We also proposed a lightweight detection head, called SDECD (Shared DEConv Convolutional Detection Head). The SDECD can minimize model parameters and computational load while maintaining accuracy. Experimental results show that RFS-YOLOv8n improves mean average precision (mAP)0.5(%) by 0.9% and mAP0.5:0.95(%) by 1.0% compared to the original model. The params was reduced by 34.7%, and GFLOPS (G) decreases by 15.0%, with mAP0.5(%) reaching 98.2%. Therefore, the lightweight and accurate model is effective for defect inspection in electronic industry.