基于Light-YOLO集成自研EMCA注意力机制的轻量化IGBT缺陷检测网络

Light-YOLO: A Lightweight IGBT Defect Detection Network Integrating a Self-Developed EMCA Attention Mechanism

摘要:

针对绝缘栅双极型晶体管模块覆铜陶瓷衬板微米缺陷人工检测效率低、漏检率高,以及现有深度学习模型计算复杂度高、难以部署的问题,提出一种集成自研注意力机制的轻量化缺陷检测网络Light-YOLO；通过构建首个IGBT-DCB工业缺陷数据集,采用Ghost模块与深度可分离卷积对YOLOv5s模型进行轻量化重构,结合加权双向特征金字塔网络优化多尺度特征融合,并创新设计高效混合坐标注意力机制以增强对微小低对比度缺陷的感知能力；在自建数据集上的实验结果表明,所提模型参数量降至3.17M,较基准模型减少55%,平均精度均值提升至81.8%,召回率达到82.1%；实现了在显著降低模型复杂度的同时检测精度的提升,为工业产线中的高精度实时缺陷检测提供了有效的轻量化解决方案。

Abstract:

To address the problems of low efficiency and high missed detection rates in manual inspection of micrometer-scale defects on insulated gate bipolar transistor module substrates, as well as the high computational complexity and difficulty in deploying existing deep learning models, a lightweight defect detection network named Light-YOLO integrating a self-developed attention mechanism is proposed. By constructing the first industrial defect dataset for IGBT-DCB, the YOLOv5s model is lightweighted using Ghost modules and depthwise separable convolution. The multi-scale feature fusion is optimized by incorporating a weighted bidirectional feature pyramid network, and an efficient mixed coordinate attention mechanism is innovatively designed to enhance the perception of tiny, low-contrast defects. Experimental results on the self-built dataset show that the proposed model reduces the number of parameters to 3.17M, a 55% decrease compared to the baseline, while increasing the mean average precision to 81.8% and achieving a recall rate of 82.1%. This achieves an improvement in detection accuracy alongside a significant reduction in model complexity, providing an effective lightweight solution for high-precision real-time defect detection in industrial production lines.