改进YOLOv8的芯片表面微小缺陷检测算法

Improved YOLOv8-Based Algorithm for Detecting Tiny Defects on Chip Surfaces

摘要:

芯片表面缺陷通常具有尺寸微小、形态复杂等特点，易导致检测过程中出现误检与漏检；针对上述问题，提出了一种基于改进YOLOv8的芯片表面微小缺陷检测算法；设计轻量级 C2f_WT 模块，扩大模型在特征提取阶段的感受野，增强对不同缺陷特征的感知能力；重新设计主干网络构建多尺度边缘增强模块，提升局部特征表达能力；重构颈部网络，增设空间与通道特征融合重组模块，实现对冗余特征的压缩与有效信息表达的增强；采用动态上采样模块，实现高效特征重映射与聚合；实验结果表明，改进后模型与YOLOv8n基线模型相比，mAP@50和mAP@50:95分别提升4.2 %和 3.8 %，参数量减少8.30 %，计算量降低25.93 %，证实了该方法在提高芯片表面缺陷检测精度与效率的同时，增强了其在实际应用中的适用性。

Abstract:

Chip surface defects usually have very small size and complex morphology, which easily cause false detection and missed detection in the inspection process. To solve these problems, an improved YOLOv8-based algorithm for chip surface tiny defect detection is proposed. A lightweight C2f_WT module is designed to enlarge the receptive field in the feature extraction stage and to enhance the perception ability of different defect features. The backbone network is redesigned by introducing a multi-scale edge enhancement module to improve local feature expression. In the neck network, a spatial-channel feature fusion reconstruction module is constructed to compress redundant features and to strengthen effective information expression. A dynamic upsampling module is adopted to realize efficient feature remapping and aggregation. Experimental results show that, compared with the YOLOv8n baseline model, the improved model increases mAP@50 and mAP@50:95 by 4.2 % and 3.8 %, respectively, reduces the number of parameters by 8.30 %, and decreases the computation by 25.93 %. These results prove that the proposed method improves the accuracy and efficiency of chip surface defect detection, and enhances its applicability in practical applications.