Research on Underwater Target Recognition Technology Based on Neural Network

Abstract

At present, the underwater environment required by the seafood aquaculture industry is very bad, and the fishing operation is completed artificially. In this environment, the use of machine fishing instead of artificial fishing is the development trend in the future. By comparing the characteristics of different algorithms, the multiscale Retinex algorithm (autoMSRCR) is selected to deal with image color skew, blur, atomization, and other problems. Labelimg software is used to annotate underwater targets in the image and make data sets. Of these, 20% are used as test sets, 70% as training sets, and 10% as verification sets. The target detection network of You Only Look Once Version4 (YOLOv4) based on convolutional neural networks (CNN) is adopted in this paper. The main feature extraction network adopts CSPDarknet53 structure, and the feature fusion network adopts SSP, and PANet network carries out sampling and convolution operations. The prediction output of extracted features is carried out through YoloHead network. After training the recognition model of the training sets, the detection effect is obtained by testing the data of the test sets. The identification accuracy of sea cucumber and sea urchin is 90.8% and 87.76%, respectively. Experiments show that the target detection network model can accurately identify the specified underwater organisms in the underwater environment.