改进YOLOv7的晶圆字符检测算法

Improved YOLOv7 Wafer Character Detection Algorithm

针对晶圆加工中的字符检测问题,提出一种基于YOLOv7改进的目标检测模型。在原版YOLOv7的SPP层之前插入Swin Transformer模块,增强网络对于全局信息的获取能力,提升对于全局和局部特征的整合能力;在预测部分插入A2-Net注意力机制,将特征信息全局融合后重新分配,提升网络的鲁棒性;在定位损失函数上用SIOU损失函数代替CIOU,角度损失的引入,增加了对于字符检测位置的准确性。在自制的字符数据集上,实验验证改进后的模型相比于传统模型,mAP提升了5.02%,并且每秒识别图片数高于传统算法,在实际使用中也取得了良好的效果。

For character detection in wafer processing,an improved object detection model based on YOLOv7.A Swin Transformer module is inserted before the SPP layer of the original YOLOv7 to enhance the network s ability to obtain global information and improve the integration ability of global and local features.Then the A2-Net attention mechanism is inserted in the prediction part,and the feature information is reallocated after global fusion to improve the robustness of the network.Finally,the SIOU Loss function is used to replace CIOU in the location loss function.The introduction of angle loss increases the accuracy of character location detection.On the self-made character dataset,the experimental results show that compared with the traditional model,the improved model improves the mAP by 5.02%,and the FPS is higher than that of the traditional algorithm.It also achieves good results in practical use.