Intelligent Recognition Technology For Drainage Network Defects Based On Improved YOLO V5

Drainage pipes are important urban infrastructure.With the aging of drainage networks,various pipeline defects have become increasingly common.Timely detection and repair of drainage pipes based on inspection results is crucial for good operation of urban drainage pipes.Closed Circuit Television(CCTV)inspection is a commonly used underground infrastructure detection technology,but it requires manual inspection of inspection videos,which is time-consuming,labor-intensive,and has poor consistency in reading results.Convolutional neural networks(CNNs)based on object detection have great potential to automatically detect defects and improve detection efficiency.The YOLO v5 model,a classic object detection algorithm,has received attention in this field due to its advantages of small scale and low deployment cost.However,in drainage pipe defect images,there are usually many small targets,so the accuracy of YOLO v5 may be affected by false detections.In order to solve this problem,this study proposes an improved YOLO v5 model to enhance the detection ability of YOLO v5 in drainage pipe defects,mainly carrying out the following work:(1)Based on domestic and foreign intelligent defect interpretation methods,especially in Western countries,deep learning models for defect classification,grading,and positioning were analyzed and studied.Combined with the research status of urban drainage pipe defect detection in China and the characteristics of defects,the classic object detection algorithm YOLO v5 model was studied and a benchmark detection process and model were constructed.(2)In view of the lack of sample libraries and uniform standards for sample library construction in the field of drainage pipe defect interpretation,the construction of a standard sample library based on the Huai’an project’s pipeline survey data was studied.Firstly,a set of standard processes for producing high-quality sample libraries was formed by analyzing the production process of other publicly available datasets in other fields and foreign drainage pipe datasets,combined with the characteristics of drainage pipe defects in China.Secondly,a set of standards for evaluating the quality of defect sample images was formed based on the image features of defects and the requirements of deep learning models.Finally,using the pipeline survey data of 21,825 pipeline sections in Huai’an,1,800 clear and classified drainage pipe defect datasets with distinct characteristics were formed after frame selection.(3)In view of the current situation that research on drainage pipe defects mainly focuses on classification and less on positioning and real-time monitoring,this paper used the bestperforming object detection algorithm YOLO v5 as the benchmark model to study the recognition of six common defects.On this basis,three improvements were made to YOLO v5: first,loss functions based on GIo U,EIo U,Focal EIo U,and SIo U were proposed,effectively improving the model’s performance;second,to improve the model’s detection effect on small targets and reduce false detections and missed detections,the shallow feature map in the backbone network was fused into the feature fusion network to add a prediction layer for small targets;finally,to improve the model’s sensitivity to regions of interest in images and reduce the interference of redundant background information,an attention module was introduced into YOLO v5.These three improvements increased the average accuracy m AP value by 2.0%,2.9%,and 5.9%,respectively.(4)The four effective improvements were integrated,and the results showed that the improved YOLO v5 model provided in this paper increased by 6.5 percentage points compared to the original model,reaching an m AP of 92.1%.This shows that the improvements made in this paper effectively enhance the detection ability of YOLO v5 for drainage pipe defects,and the detection speed of this model can support direct detection of videos.