结合Transformer和SimAM轻量化路面损伤检测算法  

Combining Transformer and SimAM lightweight pavement damage detection algorithms

作　　者：杨杰 蒋严宣 熊欣燕 YANG Jie;JIANG Yanxuan;XIONG Xinyan(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000

出　　处：《铁道科学与工程学报》2024年第9期3911-3920,共10页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(20191258);国家“863”计划资助项目(2007AA11Z180);国家重点研发计划项目(2023YFB4302100);江西省重大科技研发专项项目(20232ACE01011);中国科学院赣江创新研究院自主部署项目(E255J001)。

摘　　要：道路表面的损坏不仅会严重影响驾驶舒适性,还对行车安全造成威胁。若不及时检测和修复,可能导致交通事故的发生。因此,及时检测路面损伤对路面安全和维护有重要意义。针对现有路面损伤检测模型中存在识别精确度低和计算量大的问题,提出一种结合Transformer和SimAM轻量化路面损伤检测算法。首先,结合Transformer的优势,在模型中引入COT模块加强特征提取性能,其可以利用特征图的上下文信息构建自注意力机制,有效捕获路面损伤图像的上下文信息,加强信息表征能力。其次,针对不同大小的路面缺陷,提出MSC模块捕获全局信息,其可以结合多个池化操作动态地增加感受野的大小。同时,MSC模块与COT模块相结合,不仅有效减少了模型的计算量和参数量,而且进一步提升了检测精度。随后,融入SimAM注意力机制调节特征,增强模型在复杂场景下的特征表达能力,抑制无关特征的干扰。研究结果表明,改进算法的平均准确率为70.1%,其精度与YOLOv7、YOLOv7-tiny、YOLOv6-m、YOLOv5-l相比,分别提升2.8%、10.9%、10%、1.4%。此外,所提模型的计算量为40.3 G,约为YOLOv7、YOLOv7-x、YOLOv6-m、YOLOv5-l的38.4%、21.4%、49%、35.2%。通过与主流目标检测模型相比,所提出的模型在提高检测精度的同时,兼顾了模型的计算复杂度,在公开数据集上取得了良好的识别效果,能够有效地检测路面损伤。Damage to the road surface can not only seriously affect driving comfort but also pose a threat to driving safety.If not tested and repaired promptly,it may lead to traffic accidents.Timely detection of road damage is important for road safety and maintenance.Aiming at the problems of low recognition accuracy and large computation in the existing pavement damage detection model,a lightweight pavement damage detection algorithm combining Transformer and SimAM was proposed.First,combining the advantages of Transformer,the COT module was introduced into the model to strengthen the performance of feature extraction,which could utilize the contextual information of the feature map to construct a self-attention mechanism to effectively capture the contextual information of the pavement damage image and strengthen the information characterization capability.Second,for pavement defects of different sizes,the MSC module was proposed to capture the global information,which could be combined with multiple pooling operations to dynamically increase the size of the receptive field.Meanwhile,the MSC module was combined with the COT module,which not only effectively reduces the computational and parametric quantities of the model,but also further improves the detection accuracy.Subsequently,the SimAM attention mechanism was incorporated to regulate the features,enhance the feature expression ability of the model in complex scenes,and suppress the interference of irrelevant features.The results show that the average accuracy of the improved algorithm is 70.1%,and its accuracy is 2.8%,10.9%,10%,and 1.4%higher compared to YOLOv7,YOLOv7-tiny,YOLOv6-m,and YOLOv5-l,respectively.In addition,the computation of the proposed model is 40.3G,which is about 38.4%,21.4%,49%,and 35.2%of that of YOLOv7,YOLOv7-x,YOLOv6-m,and YOLOv5-l.By comparing with mainstream target detection models,the proposed model can take into account the computational complexity of the model while improving the detection accuracy,and achieve good recognition results on

关 键 词：路面损伤检测 YOLOv7 卷积神经网络 TRANSFORMER SimAM 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]