Wafer Defect Detection Based on YOLO-BA

With the advancement of the chip industry, manufacturing processes have become increasingly sophisticated, resulting in a significant rise in the scale and complexity of integrated circuits. This complexity has rendered wafer manufacturing processes more intricate, leading to a higher probability and diversity of wafer defects. To enhance production yield and improve process control, it is essential to identify defects and address the corresponding issues. This paper presents an improved wafer detection algorithm, YOLOBA, based on YOLOv10n. The approach incorporates Bottleneck Transformers (BoT) and Adaptive Kernel Convolution (AKConv) into the backbone network, effectively suppressing interference from complex background information while reducing feature redundancy in both spatial and channel dimensions. This facilitates more efficient feature extraction of wafer defects while maintaining the lightweight nature of the model. Experimental results demonstrate that the improved algorithm achieves a 0.9% higher recall rate, a 0.8% increase in [email protected]_0.95, and a 0.4% improvement in [email protected] compared to the latest YOLOv10n model, all while maintaining similar parameter counts and GFLOPs.