【目的】高功率半导体激光器的光学灾变损伤是限制其可靠性和寿命的主要因素,因此,有效的缺陷检测对于优化激光器芯片的制造工艺和结构设计至关重要。提出了一种基于改进YOLOv7-Tiny的轻量化激光器芯片缺陷检测算法,旨在解决深度学习应用于缺陷检测时面临的高计算量和参数量问题。【方法】利用轻量化卷积神经网络替换特征提取主干有效减少对计算资源消耗,有效提取电致发光图像中缺陷特征。为从上下文特征获取更丰富的信息,引入多分支重参数化卷积块重构聚合模块,通过多路径分支丰富特征表示,训练与推理的解耦保证检测效率。此外,结合坐标注意力,提升定位精度。进行了剪枝实验和模型部署,验证算法的初步应用。【结果】在电致发光缺陷数据集上的实验结果显示,本文方法能在较低的参数和计算量下准确地检测出芯片缺陷,展现出良好的性能。

【Purposes】Catastrophic Optical Damage (COD) is a major limiting factor for the reli‐ ability and lifespan of high-power semiconductor lasers, making effective defect detection crucial for  optimizing the manufacturing processes and structural designs of laser chips. In this study, a light‐ weight laser chip defect detection algorithm based on an improved YOLOv7-Tiny is proposed, aim‐ ing at addressing the high computational and parameter demands of deep learning applications in defect  detection. 【Methods】 By employing a lightweight convolutional neural network as the feature extrac‐ tion backbone and integrating multi-branch reparameterized convolution blocks, this algorithm not  only significantly reduces resource consumption but also enhances feature representation capabilities.  Additionally, the introduced coordinate attention mechanism improves the precision of defect localiza‐ tion. Pruning experiments and model deployment are conducted to further verify the algorithm practi‐ cality. 【Findings】 Experimental results on the electroluminescence dataset demonstrate that this  method can accurately detect chip defects with lower parameter and computational costs, showing ex‐ cellent performance.