基于改进Yolov5的晶振多缺陷检测

Multi-defect detection of crystal oscillators based on improved Yolov5

长期以来,晶振的缺陷检测主要依靠人工,存在着效率低、受外界影响程度大、成本高昂的问题。传统方法无法实现多缺陷一次性记录检出,而深度学习应用部署于工业界也存在着无GPU加速导致的推理速度慢的问题。针对以上问题,选定算法模型并通过在输入模块对图像增强、主干网络特征图融合、颈部网络引入新的尺度检测层和新的激活函数、输出预测时引入注意力机制的多种方式进行改进,在保留模型轻量化的同时增加小目标检测能力。最后进行模型剪枝和硬件加速推理,部署于PC端进行缺陷统计等工作。实验表明该系统的检出率在95%以上,实现了晶振缺陷检测的自动化和精细化。

The detection of defects in crystal oscillators has mainly relied on manual work,which has the problems of low efficiency,high degrees of external influence and high cost.Traditional methods are unable to detect multiple defects in a single record,while deep learning application in the industry also suffers from slow inference due to a lack of GPU acceleration.To address these issues,the algorithm model is selected to improve the detection in various ways,including conduction of image enhancement in the input module,feature map fusion of the backbone network,introduction of a new scale detection layer and new activation functions into the neck network,and addition of the attention mechanism into the output prediction section.These efforts increase the tiny target detection capability while retaining the lightweight of the model.Finally,model pruning and hardware-accelerated inference are carried out,and the research is deployed on a PC to perform defect statistics and other tasks.The experiments show that the detection rate of the system is above 95%,realizing the automation and refinement of defect detection of crystal oscillators.