Research On Building Crack Detection Based On YOLOv5 Algorithm

With the rapid development of the construction industry,the detection and identification of building cracks have become an important task,but the traditional detection methods are difficult to detect,high false detection rate,high omission rate and low efficiency.As an important part of computing and visual field,YOLO series algorithm has the advantages of better detection performance and more stable detection effect.At present,YOLOv5 algorithm has occupied a dominant position in the field of target detection with its excellent detection effect.However,the detection effect is not the same in different scenarios.It is necessary to look at the requirements for target detection in different scenarios,and then optimize and improve the YOLOv5 target detection algorithm.The experimental results show that the performance(including speed and recognition accuracy)of the network model can be improved by adding appropriate attention mechanism to the YOLOv5 algorithm.According to the current mainstream and relatively mature YOLOv5 target detection algorithm,this paper chooses building cracks as the research target and conducts corresponding research on its detection problem.The main work is as follows:(1)For the reasons of insufficient detection and high false detection rate in the current building crack detection process,this paper proposes to add SE attention mechanism,CBAM attention mechanism and ECA attention mechanism to YOLOv5 target detection algorithm.The experimental results show that the performance(including speed and recognition accuracy)of the network model can be improved by adding appropriate attention mechanism to the YOLOv5 algorithm.Adding attention mechanism to target detection algorithm improves the channel and spatial relationship of target features.It enhances the network’s attention to the key information in the feature map,which is conducive to the network’s more perfect extraction and utilization of features.The detection network can extract important information according to the weight,which increases the detection ability of the network.(2)Mosaic data enhancement method was used to enrich the data set,avoid overfitting in training,and enhance the robustness of the network model.Finally,three improved YOLOv5 network models were obtained by adding the improved method of attention mechanism,and the optimal model was obtained by comparison: ECAYOLOv5.It effectively enhances the overall performance of network detection and reduces the false detection and missing detection.(3)Based on the improvement of(2),it is found that the network model has too much data,resulting in slow running time.Therefore,a lightweight network model of YOLOv5 is proposed,which includes shufflenetv2 and mobilenetv3-small.Through comparison,it is found that the mobilenetv3-small network model ensures the detection accuracy,reduces the computational complexity and improves the detection speed.Then the improved model is used to test the building cracks,and the results show that the network model can meet the working requirements.