The Research On Detection Technology Of Floating Objects On Water Surface Based On Deep Learning

With the increasingly severe water surface environmental pollution,how to maintain the water surface clean has always been a difficult problem.If the floating garbage on the water surface is not cleaned up in time,it will have a serious impact on people’s daily life.If the manual salvage method is adopted,not only the safety factor is low,but also the efficiency is low.Nowadays,with the advent of the era of big data and the rapid rise of artificial intelligence technology,various kinds of intelligent products emerge at the historic moment,including water cleaning robots.Target detection,as one of its key technologies,it is also very necessary to study them.The application of deep learning to the field of image recognition is often accompanied by a large number of computing quantity and parameters,so it is difficult to deploy the model to the mobile terminal.So this paper studies a few popular object detection algorithms.(1)For YOLOv4 target detection algorithm network parameter number and calculation,high requirements for embedded equipment performance,using deep separation method instead of 33 conventional convolution in the network,at the same time using MobileNet series of three lightweight network respectively as the backbone of YOLOv4 network to replace the original CSPDarknet53 way to realize the lightweight of the whole model.The lightweight improved models are compared with the original model on the surface floating object dataset,and the results show that the overall performance of the MobileNetV3 improved model is the best,and the detection accuracy is the smallest difference with the original model,but the number of parameters and calculation amount are only 1/5 of the original model,and the detection speed is 1.7 times higher.(2)In view of the slight decrease in the detection accuracy of the lightweight improved model,the SFPN is adopted in the neck network of the model.SFPN structure is in the characteristics of the original FPN structure on the basis of linear scaling to transform two adjacent feature layer to the same scale,and then transform features through the add additive synthesis,get a new scale of synthetic layer,the new synthesis layer scale between the original two adjacent feature layer scale.The model adds new feature layers through this synthetic fusion way,making the feature fusion between the layers smoother.Through the comparison experiment of the model,we conclude that the synthetic fusion pyramid structure can improve the detection accuracy of the lightweight model with the same parameters and the original YOLOv4 model,but the parameters and calculation amount of the model are much lower than the original model,which greatly improves the detection speed of the model.Through deep learning-based object detection technology,a lightweight YOLOv4 object detection method for real-time detection of floating objects.