Deep Neural Network Design And Acceleration For Small-sample Underwater Target Recognition

With the wide application of sonar technology in national defense construction,ocean exploration and other fields,the problem of underwater target recognition has gradually become an important research content of the intelligent process of sonar.Due to the difficulty and high cost of acquiring underwater target data,the underwater sample size is scarce,and it is difficult to meet the current data requirements of machine learning,especially deep learning.Therefore,a set of small sample numbers with high accuracy and fast recognition speed have been developed.The underwater target recognition system has important practical significance.Based on the Siamese network architecture and deep convolutional neural network,this paper designs a network model for measuring the similarity of underwater target samples.This article uses actual samples and simulated virtual samples to train the model,and calculates the similarity between the two samples to be tested through the trained model,and then determines whether the two samples are the same type of target,which effectively solves the difficulty to identify small-sample underwater targets.First of all,this paper conducts in-depth research on the modeling and feature extraction methods of underwater target radiated noise.Based on real data,7 types of underwater target samples are designed for network model training.Secondly,in order to evaluate the generalization ability of the designed network model,based on the above 7 types of underwater target samples,samples with different Doppler frequency offset samples,different signal-to-noise ratios and different numbers of interference spectrum lines are designed as the evaluation data of the network model.Finally,in order to test the recognition ability of the network model,the obtained network model is tested using real underwater target samples.The test results show that when the Doppler frequency deviation,signal-to-noise ratio and interference spectrum number are within a certain range,the network model in this paper can recognize the test sample with an accuracy of over85%.In addition,this article is based on FPGA to realize the optimization and acceleration of the deep neural network,and comprehensively considers the resource utilization and recognition rate and other limiting factors,and designs an embedded small sample underwater target recognition system.This system is conducive to promoting the engineering application of small sample underwater target recognition technology.