基于PNC-YOLOv8的惯导PCB缺陷检测研究

Research on defect detection of inertial navigation PCB based on PNC-YOLOv8

摘要:

针对惯导维修中线路板缺陷检测复杂度高、精度低以及效率低下的问题,提出一种改进的YOLOv8算法PNC-YOLOv8。该算法首先引入高分辨率小目标检测层构建多尺度特征金字塔,有效增强微小缺陷的特征表达能力；其次,在主干网络嵌入新型NAMAttention注意力机制,通过建立跨通道长程依赖关系提升复杂背景下的特征判别能力；最后,在特征融合阶段集成CBAM注意力模块,通过通道-空间协同注意力机制实现局部关键特征的强化。在数据集PKU-Market-PCB上的结果表明,PNC-YOLOv8网络平均精度达到93.8%,与基础网络YOLOv8相比,平均精度提高了3.9%,与其他主流检测网络相比,性能也有显著提升,同时其单帧检测速度达到了100ms。PNC-YOLOv8网络为惯导维修中线路板缺陷检测提供了新的方案,具有一定的应用价值。

Abstract:

In order to solve the problems of high complexity, low accuracy and low efficiency of circuit board defect detection in inertial navigation maintenance, an improved YOLOv8 algorithm PNC-YOLOv8 was proposed. Firstly, a high-resolution small target detection layer is introduced to construct a multi-scale feature pyramid, which effectively enhances the feature expression ability of small defects. Secondly, a new NAMAttention attention mechanism is embedded in the backbone network to improve the feature discrimination ability in complex backgrounds by establishing cross-channel long-range dependencies. Finally, the CBAM attention module was integrated in the feature fusion stage, and the reinforcement of local key features was realized through the channel-space collaborative attention mechanism. The results on the dataset PKU-Market-PCB show that the average accuracy of PNC-YOLOv8 network reaches 93.8%, which is 3.9% higher than that of the basic network YOLOv8, and the performance is also significantly improved compared with other mainstream detection networks, and its single-frame detection speed reaches 100ms. The PNC-YOLOv8 network provides a new solution for circuit board defect detection in inertial navigation maintenance, which has certain application value.