Research And Application Of Ship Target Detection Technology Based On Lightweight Network

In recent years,with the development of deep learning software and hardware technology,artificial intelligence equipment based on deep learning is becoming more and more mature,including unmanned surface vehicle which is often used to detect,track and measure targets and obstacles on the sea.The marine target detection based on computer vision is an important sensor unit to assist the unmanned surface vehicle to move.It is of great significance to study the marine ship object detection algorithm.The development of deep learning technology makes the detection speed and accuracy of various target detection algorithms reach a very high level with the support of high-performance GPU devices.This is mainly due to the complex network structure and massive training data of these algorithms.However,at the same time,it is accompanied by the huge amount of parameters and calculation of network model and the requirement of high power consumption and high computing power of hardware devices,which leads to the following problems The deployment of the model to low configuration platform such as mobile or embedded is limited.Therefore,how to use as little computational resources as possible and achieve the high performance of target detection model,that is,lightweight network design,has become the focus of academic research.Aiming at the above problems,this paper studies the ship target detection technology based on lightweight network structure,which can ensure the accuracy of ship target detection,reduce the model volume and calculation as much as possible,and improve the detection speed.Through the comparative analysis of common target detection algorithms,the one-stage target detection algorithm YOLOV3 is selected as the main research target,and the lightweight improvement of the model is mainly carried out from the two aspects of basic backbone network and deep feature fusion network.In the improvement of the basic backbone network,this paper uses the idea of MobileNetV2 and ShuffleNetV2,which have excellent lightweight performance,to simplify and optimize the backbone network of YOLOV3.Combined with the characteristics of lightweight target detection task,different improvement strategies are designed,and the algorithm training and evaluation are realized on the server.Through the comparative analysis of the experimental results on the self built ship target data set,the algorithm is selected Among them,the model structure performs better.In the improvement of deep feature fusion network,this paper will mainly design different improvement methods for convolution neural network,BN layer and activation function in its unit structure,and verify the effectiveness of each method through experiments to complete the preliminary design of lightweight ship target detection model.On the basis of the above designed model,the paper introduces pruning,quantization,distillation and other model compression and acceleration methods to further improve the overall lightweight of the model,so as to obtain smaller model size and faster detection speed,and reduce the accuracy loss in the improvement as much as possible.The specific effect of the model lightweight improvement is analyzed by comparing the model training results with the original model performance of YOLOV3.In this paper,the lightweight ship target detection model verified by the experiment is translated into the model file suitable for the embedded environment of unmanned surface vehicle through the compressed model deployment tool,and the converted target detection model is simulated in the corresponding CPU environment.The real-time detection effect of the model in the actual marine environment is explored through the experimental results.