Research On Deep Learning-based Underwater Target Image Recognition Method

At present,there are few methods for underwater target recognition based on underwater sonar images.The main reason is that it difficult to extract the features of underwater target sonar image manually,which will result in the problem of low target recognition rate.However,Deep learning is effective in the field of natural image recognition and target detection.The deep learning method can automatically extract the underwater target sonar image features,which may avoid the problem of information loss and greatly improve the target recognition rate.The recognition object in this paper is mainly aimed at small underwater targets,using underwater sonar targets images as data sets,and exploring new application f of deep learning methods in the field of underwater target image recognition.In the aspect of data preprocessing,the original sonar image data has been denoised,the data set for feature extraction network training and target detection and recognition network training is be made.During the model establishment,considering the application requirements of actual projects,the YOLOv3 target detection and recognition model is selected as the reference model.Then the network structure and strategy details are analyzed and discussed.The advantages of target recognition and the specific application scenarios of this paper are improved and simplified.Then the corresponding joint loss function is designed to unify the target detection and target recognition tasks.Then the nonlinear activation function,regularization method and parameter optimization method are analyzed and compared in detail and the appropriate strategy is selected.Finally,this paper tests and evaluates the performance and robustness of the feature extraction network from TPR,FRP,ROC,AUC and other evaluation indicators,from the average recognition accuracy,IOU and model identification time to the entire underwater sonar.The performance and robustness of image target detection and recognition networks were tested and evaluated.Through the simulation experiment,the underwater sonar image target feature extraction network designed by this paper performs well.In the noise-free test set,the recognition rate of pipeline,frogman,air,fish group and propeller has more than 98%.As for the robustness of the network,the network is tested and the ROC and AUC are measured by the noise-added data set.The test results show that the recognition accuracy of the five categories is greater than 80% when the average signal-to-noise ratio is-9.0dB.The values are all greater than 0.8;and the overall target detection and recognition network,in the VOC2007 format test set,the average recognition accuracy of the five different targets is 79%,the average IOU value is 0.92,the model performance is good.Finally,the whole model was migrated to the embedded platform and developed into a complete underwater target recognition system.The system operation results were consistent with the simulation results,and the average recognition speed was 28 FPS,which basically met the real-time requirements.