Research On Ship Target Classification Technology Based On Deep Learning

With the steady progress of China’s "maritime power" strategy,marine security has gradually become a hot issue.It is an important issue to ensure marine security to truly know each other in the ocean and accurately and efficiently classify and identify ship targets.The advancement of research must be accompanied by the progress of science and technology.The rapid development of artificial intelligence has just laid a solid technical support for the automatic classification and recognition of ship targets.With the arrival of a new wave of scientific and technological revolution,the field of ship target classification and recognition will also have significant development.This paper focuses on the central task of ship target classification,and takes the composition of ship target classification and recognition system as the design principle to study the practicability of deep learning method in this task field.The specific research contents are as follows:1.Based on the generation mechanism of ship radiated noise,the modeling,simulation and feature extraction of ship radiated noise are carried out in this paper.From the level of physical parameters and spectral characteristics,the mathematical statistical model is used to model and simulate the ship target radiated noise according to the basic parameter settings of the experimental data set.Then,LOFAR spectral analysis method and Mel cepstrum analysis method are used to extract features of the simulated ship target radiated noise.From the principle of generating radiated noise,different types of ship targets have differences and separability in different characteristics,and line spectrum characteristics are the stable characteristics of ship targets.2.Using ship radiated noise characteristics as the theoretical basis for ship target classification,this paper studies the application of SVM,LSTM,Res Net and Vi T in the field of ship target classification.The first two methods are mainly used for the processing of discrete data series constructed by Mel cepstrum coefficients,and the continuous spectrum and line spectrum features are mixed.However,in the multi classification of ship targets,with the increase of continuous spectrum features,the classification effect is not ideal.The latter two methods are mainly used in the processing of the image data set constructed by LOFAR spectrum.By observing the visualization process of the thermogram,it can be obtained that the stable features in the features extracted by the neural network are mainly concentrated on the online spectrum.Therefore,on the basis of LOFAR spectrum,this paper strengthens the line spectrum features,making the final classification effect of the two methods better.3.Based on the application research of the above four methods,this paper proposes an improved hybrid Vi T network structure,which integrates multi head attention mechanism and convolution operation,and improves the practicability of the network in the field of ship target classification.The multi head attention mechanism can smooth the loss domain and improve the performance of the model,while the convolution operation can more fully perform feature extraction.This paper proves the complementarity between attention mechanism and convolution operation through experiments,and proves the feasibility of network improvement.Simulation and experimental data verify the effectiveness of the proposed algorithm,and the experimental results show that the proposed algorithm in this paper is superior to existing algorithms in classification accuracy and robustness.