融合改进YOLOv7与UNet的编码点定位方法  

Fusion of improved YOLOv7 and UNet codepoint localization methods

作　　者：刘升[1] 古丽巴哈尔·托乎提[1,2] 补生来 买买提明·艾尼 Liu Sheng;Gulbahar Tohti;Bu Shenglai;Mamtimin Geni(College of Intelligent Manufacturing Modern Industry,Xinjiang University,Urumqi 830047,China;State Key Laboratory for Manufacturing Systems Engineering,Xi′an Jiaotong University,Xi′an 710049,China)

机构地区：[1]新疆大学智能制造现代产业学院,乌鲁木齐830047 [2]西安交通大学机械制造系统工程国家重点实验室,西安710049

出　　处：《国外电子测量技术》2024年第5期9-17,共9页Foreign Electronic Measurement Technology

基　　金：国家自然科学基金地区基金(12162031);西安交通大学机械制造系统工程国家重点实验室(sklms2022022)项目资助。

摘　　要：针对编码点的定位存在误检率高和精准度差的问题,提出了一种融合改进YOLOv7与UNet的圆形编码点定位方法。第1阶段使用改进的YOLOv7检测编码点的位置,改进的YOLOv7首先将DCN-v2可变形卷积引入ELAN模块,提升特征提取能力;其次把卷积块注意力模块(CBAM)机制嵌入骨干网络使网络更关注目标特征;然后使用Focal-EIoU Loss提高收敛速度;最后构建OD-Cat模块替换ConCat模块以提升网络检测精度。提取出每个圆形编码点的ROI后,第2阶段通过UNet分割出编码点的中心轮廓后,使用最小二乘法拟合出编码点的中心。实验结果表明,改进后的模型比原YOLOv7的精确率提高了6.33%,平均精度均值(mAP)提升了5.76%;提出的定位方法验证了在噪声、亮度不足或曝光等复杂环境下可以准确定位出编码点的中心椭圆轮廓,在实际工业视觉测量中具备鲁棒性。Aiming at the problems of high misdetection rate and poor accuracy in the localization of codepoints,a circular codepoint localization method fusing improved YOLOv7 and UNet is proposed.In the first stage,the improved YOLOv7 is used to detect the location of coding points.The improved YOLOv7 firstly introduces DCN-v2 deformable convolution into the ELAN module to improve the feature extraction ability.Secondly,the CBAM attention mechanism is embedded into the backbone network to make the network pay more attention to the target features.Then,Focal-EIoU loss is used to improve the convergence speed.Finally,OD-Cat is constructed to replace the ConCat module to improve the network detection accuracy.Module to replace the ConCat module to improve the network detection accuracy.After extracting the ROI of each circular coding point,the center contour of the coding point is segmented by UNet in the second stage,and then the center of the coding point is fitted using the least squares method.The experimental results show that the improved model improves the precision by 6.33%and the mean average precision(mAP)by 5.76%over the original YOLOv7.The proposed localisation method verifies that it can accurately locate the central ellipse contour of the coded point under complex environments such as noise,insufficient brightness or exposure,and is robust in practical industrial vision measurements.

关 键 词：编码点识别 深度学习 中心定位 YOLOv7 UNet 最小二乘法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]