Research And Application Of River Floating Garbage Detection Based On YOLOv5s

In recent years,with the implementation of the "River Chief System" management system,the phenomenon of factories’ centralized pollutant discharge into rivers has basically disappeared,and the water environment has been significantly improved.However,there are still littering by residents,and some domestic garbage floating on the water.Traditional manual inspections,recording pollutant distribution locations,and reporting river monitoring methods are inefficient,costly,and lack information transparency;At present,automatic inspection methods based on machine vision only use target detection algorithms to identify floating garbage,and rarely perform geographic location and pollution degree analysis on detected garbage,which is not of practical value.To this end,this paper designs a river floating garbage detection model based on improved YOLOv5 s,which uses deep learning target detection algorithms to detect garbage on river drone images,and visualizes the detection results on the Web GIS platform and provides corresponding statistical information of river garbage,To improve the work efficiency and information transparency of the river supervision system.The specific research work of this paper is as follows:(1)In order to realize the classification and detection of floating garbage in the river with different pollution levels,and extract the geographic location information of the image to quickly locate the garbage,and clarify the pollution degree of different sections of the river.Based on the RTK-UAV images of multiple rivers in Xuzhou City,this paper classifies the garbage in the images according to the degree of pollution and makes a small river floating garbage data set.(2)In view of the shortcomings of the YOLOv5 s target detection algorithm in detecting small targets and unbalanced targets,the K-means clustering method is used to recalibrate the prior anchor frame of the algorithm,and the data enhancement,SPP and loss function modules are targeted.Sexual improvement.Among them,a 3×3convolution kernel is added to the SPP module to further improve the receptive field of the model,thereby improving the detection accuracy of small targets;In the loss function module,instead of the binary cross entropy function,the Focal function is used to calculate the confidence and category loss to solve the problem of imbalance between positive and negative samples.Through the performance test of the algorithm before and after the improvement in the self-made data set,it is found that the improved algorithm reduces the missed detection rate of small targets,the accuracy of class equalization is increased by 3.47%,and the detection accuracy of different categories is also available.In addition,in order to make the garbage detection information generated by the algorithm have the ability to be visualized on the electronic map,a module for reading the longitude and latitude of the image center point and the flying height of the drone,the image collection time,the image width and height information is added to the algorithm.(3)Based on the trained and improved YOLOv5 s detection algorithm,the Koa framework in the node.js platform is used to establish a visualization system for the degree of river pollution.The statistical analysis module of the system uses the principle of small hole imaging and UTM projection to determine the range of the river covered by the orthophoto,uses the analytic hierarchy process to quantify the degree of pollution in each section,and stores the statistical analysis results in Geo JSON format;The front end of the system adopts the Cesium map engine framework,and realizes the function of visualizing the pollution degree of each section of the river on the electronic map of the Web by analyzing the Geo JSON file generated by the statistical analysis module.This paper has 47 pictures,10 tables and 80 references.