基于改进YOLOv7的钢材表面缺陷检测方法

Steel surface defect detection method based on improved YOLOv7

为解决工业钢材表面缺陷检测过程中存在的检测精度低、效率低的问题,文章对原有的YOLOv7算法进行了改进,从而形成了YOLOv7-M算法。首先,在主干网络后加入GAM模块,以增强模型的全局交互能力。然后,在YOLOv7网络的预测头前端添加集成Transformer和卷积的轻量级模块MobileNetViTv3block,来捕获全局细节信息,同时,降低YOLOv7的模型参数,以提高检测速度。最后,在定位损失函数上用SIoU-Loss函数代替CIOU函数,以提升钢材缺陷检测位置的准确性。以NEU-DET数据集为基础做相关的消融与比较实验,在VOC2012数据集上做通用性的比较实验。结果表明:经过改进后的YOLOv7-M算法具有更大的优势,mAP值达到了90.75%,平均检测精度提高了7.89%,FPS速度提升了7,并且改进后的算法具有通用性。

In order to solve the problems of low accuracy and low efficiency of multi-category defect detection in the process of surface defect detection of industrial steel,the original YOLOv7 algorithm is improved in this study,and a new algorithm YOLOv7-M is formed.Firstly,the GAM module is added to the backbone network to enhance the global interaction capability of the model.Then,MobileNetViTv3 block,a lightweight module inte-grated with Transformer and convolution,is added to the prediction head of the YOLOv7 network to capture glob-al details,while reducing the model parameters of YOLOv7 to improve the detection speed.Finally,the SIOU-Loss function is used to replace the CIOU in the positioning loss function to improve the accuracy of the steel de-fect detection position.Based on the NEU-DET dataset,the relevant ablation and comparison experiments are carried out on the NEU-DET dataset,and the general comparison experiments are carried out on the VOC2012 dataset,and the results proved that the improved algorithm YOLOv7-M has greater advantages,the mAP value reaches 90.75%,the average detection accuracy is increased by 7.89%,and FPSthe speed is increased by 7.The improved algorithm is also universal.