Research On Recognition Algorithm Of Floating Objects On Water Surface Based On Deep Learning

The protection of the water environment has always been the top priority of the country’s environmental protection,and a large amount of money is invested every year to prevent water pollution.However,due to the limitations of equipment and technology,the current prevention and control methods mainly rely on manpower,such as manual river patrol and manual salvage,which are undoubtedly inefficient and unsafe.Therefore,the realization of intelligent river monitoring will undoubtedly liberate a lot of manpower and material resources.In the real scene,the monitoring methods of rivers are diverse,so the carrier requirements for algorithms are also diverse.According to different carriers,this article optimizes the algorithm from the perspective of the client and the server.The details are as follows:(1)From the perspective of the client,the current conventional deep learning model has the disadvantages of large amount of calculation and poor portability.In response to these problems,this paper proposes to use the cheap and efficient connection method in the lightweight network Ghost Net to improve the backbone network in YOLOv3.Without greatly reducing the accuracy of the original model,it is replaced by linear operations with less calculations.Partial convolution operation greatly reduces the amount of model parameters and FLOPs.Due to the problem of unbalanced categories in the data set of drifting objects collected in this subject,the trained model will not achieve the expected effect on the categories with fewer training samples.In response to this problem,this subject introduces Focal loss into the model.Focal loss function enables the model to dynamically adjust the training weight of a certain type of sample during training,thereby alleviating the problem of imbalanced categories.Comparative experiment results show that YOLOv3 using the Ghost Net connection method has an improvement of about 14% in the m AP index compared to the YOLOv3 algorithm,and the use of the Focal loss function also brings a 0.9% improvement in the algorithm’s m AP index.(2)From the perspective of the server,the server equipment is better than the client equipment in terms of hardware and software,and the general algorithm deployment server has a graphics card.Compared with the client equipment,the server has more disposable computing power.Therefore,the optimization of the server is more to improve the accuracy of the algorithm.This article is based on the dense connection method in the VOVNet network to improve the backbone network in YOLOv3.The densely connected network structure can alleviate the loss of shallow features in the deep network,thereby The features in the network extraction process can be used more efficiently.Comparative experiment results show that YOLOv3 using the VOVNet connection method improves m AP by 14.1%,and the Focal loss function brings a 1%improvement to the algorithm’s m AP indicator.Finally,based on the above-mentioned research on the client-side surface floating object detection optimal algorithm Focal-Ghost-YOLOv3,a version of the surface floating object recognition application was developed to demonstrate the value of the algorithm in the actual scene while showing the surface floating object.The results of algorithm research.