Combining Transformer and SimAM lightweight pavement damage detection algorithms
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 public datasets,which can effectively detect road damage.
万方数据
维普
