基于轻量化YOLOv7-tiny的铝材表面缺陷检测方法  

Aluminum Surface Defect Detection Method Based on Lightweight YOLOv7-tiny

作　　者：席凌飞 伊力哈木·亚尔买买提[1] XI Ling-fei;YILIHAMU·Yaermaimaiti(School of Electrical Engineering,XinJiang University,Urumqi 830017,China)

机构地区：[1]新疆大学电气工程学院,乌鲁木齐830017

出　　处：《科学技术与工程》2024年第27期11786-11794,共9页Science Technology and Engineering

基　　金：国家自然科学基金(62362063,61866037)。

摘　　要：为了提高铝材表面缺陷小目标检测效率,轻量化检测网络,提出一种基于改进YOLOv7-tiny的铝材表面缺陷检测算法。首先,在网络中加入SimAM(similarity-based attention module)注意力机制,使网络更多的注意到有用的信息,抑制图像中无效样本的干扰。其次,将主干网络中的sppcspc结构改进为Ghostsppcspc,减少的模型训练时的参数冗余,同时在检测层用GSconv代替普通卷积,轻量化网络的同时加强特征融合,提升网络检测精度,最后引入NWD(normalized wasserstein distance)结合原有的CIOU(complete intersection over union)损失函数,提升网络对小目标检测精度。将改进算法应用到天池铝材数据集中进行验证,实验结果表明,该模型能够有效识别铝型材表面不同种类的缺陷,较原YOLOv7-tiny算法mAP提高10.1%,参数量较原模型下降6.4%,计算量较原模型下降12.2%。所提方法实现了轻量化网络模型的同时,能够满足目前铝型材工厂生产现场缺陷检测要求。In order to improve the efficiency of small target detection of aluminum surface defects,an aluminum surface defect detection algorithm based on improved YOLOv7-tiny was proposed.Firstly,the SimAM(similarity-based attention module)attention mechanism was added to the network to make the network pay more attention to the useful information and suppress the interference of invalid samples in the image.Secondly,the sppcspc structure in the backbone network was improved to Ghostsppcspc,which reduced the parameter redundancy during model training,and GSconv was used in the detection layer instead of the ordinary convolution,which lightened the network and strengthened the feature fusion to improve the network detection accuracy,and finally,the NWD(normalized wasserstein distance)was introduced to replace the original CIOU(complete intersection over union)loss function,which improves the network's accuracy in detecting small targets.The improved algorithm was applied to the Tianchi aluminium dataset for validation,and the experimental results show that the model was able to identify different kinds of defects on the surface of aluminium profiles effectively,with a 10.1%increase in mAP,a 6.4%decrease in the number of parameters,and a 12.2%decrease in the amount of computation compared with the original YOLOv7-tiny algorithm.The proposed method achieves a lightweight network model and can meet the current requirements of defect detection in the production site of aluminium profile factories.

关 键 词：缺陷检测 YOLO v7-tiny 注意力机制 NWD GSconv Ghostsppcspc 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]