Research On Underwater Target Detection Method Based On Sonar Image Feature

Underwater target detection is one of the popular research contents in the field of hydroacoustics,which plays a great role in ocean exploration,underwater detection,and military information reconnaissance.In this paper,for different sonar detection equipment in underwater target detection,using sonar images and echo signals in two data forms,combined with sonar image characteristics,the study of forward-looking sonar underwater target detection,side-scan sonar underwater target detection and single-beam sonar echo signal-based underwater target classification methods,to achieve a higher accuracy detection and classification results.For the problem of serious false alarm and more missed detection in underwater small target detection by forward-looking sonar images.The multi-algorithm fusion detection of FCM,Kmeans and LCM is introduced to effectively reduce the false alarm targets in the images.Then,classical image segmentation methods such as pulse-coupled neural network and threshold segmentation are compared to obtain the target boundary contour in the target ROI,and use image shape features and geometric features to further reduce the false alarm probability and achieve accurate detection of forward-looking sonar underwater targets.For the problems of low contrast and poor effectiveness quality of side-scan sonar images,anisotropic filtering and Retinex method are introduced to make the image as a whole more consistent with human eye observation habits while preserving the image texture.Then,CFAR and its improvement methods are used to verify their effectiveness in side-scan sonar target detection,respectively.Finally,feature extraction is combined with image segmentation,and spatial and texture information are introduced into FCM and K-means,respectively,and it is experimentally verified that GFCM combined with GMRF outperforms traditional clustering methods in terms of segmentation performance as well as anti-interference performance.Also,texture features such as LBP and GLCM are proposed to be introduced into image segmentation,and the applicability of the algorithm is verified in the measured data.For the problems of inadequate feature extraction and low classification accuracy of singlebeam sonar echo signals,the use of signal two-dimensional representation is proposed for initial feature extraction of signals.The feature extraction methods with different ideas are investigated in two directions: signal two-dimensional coding and time-frequency analysis.Then,convolutional neural networks are combined to further extract features and achieve classification recognition.During the construction of deep learning network models,the necessity of migration learning methods in underwater target datasets and the ability of null convolution in improving network performance are verified.Finally,through pool experiments,different position angles of rotating targets are successfully identified and a high accuracy rate is achieved.