Study On Feature Extraction Of Underwater Moving Target Based On Acoustic Image Sequences

With the development of underwater unmanned equipment and technology,the problem of underwater security has become more and more prominent,and how to detect and identify underwater moving targets has become an urgent problem to be solved.Accurate recognition cannot be separated from the extraction of target features,among which the extraction of motion features is particularly important.Existing research in the field of underwater target recognition lacks in-depth exploration of underwater target motion features,and still remains in the stage of using simple information such as speed and path as motion features,ignoring the motion information contained in the correlation between frames in sonar image sequences.In recent years,3D moments feature extraction methods have been gradually applied to 3D image feature extraction because of its universality,convenience and spatio-temporal integration processing ability.Therefore,the 3D moments method is applied to the feature extraction of underwater moving target for the first time in this paper,so as to achieve the motion feature description for fish and underwater robot.Underwater target pattern recognition is a systematic project,which not only needs to extract features to represent the identity of the target,but also needs the cooperation of target detection,image preprocessing and other related technologies.Therefore,on the basis of fully investigating relevant literature,this paper explored a variety of target detection algorithms suitable for acoustic image sequences,and used relatively mature image preprocessing technology to process the sonar image sequences of two categories of targets acquired through self-experiment.Three motion feature extraction algorithms,3D central moments,3D velocity moments and 3D Zernike velocity moments,were used to extract the features from the processed image sequences,and the classification performance of the three features was compared and analyzed from two aspects of similarity measurement and KNN classifier recognition.A binary pattern recognition framework for underwater moving targets is established.The results show that:(1)The Gaussian mixture model simulates the changes of the noise background in the sonar image sequence over time,and can basically eliminate the interference of the background noise.The target detection algorithm based on this model extracts the foreground target with complete shape and clear outline.It is suitable for the target detection in sonar image sequences with low signal noise ratio(SNR).(2)The KNN classifier classifies the features obtained by the three feature extraction algorithms,and the recognition rate reaches more than 90%.It is proved that the three algorithms can represent the motion characteristics of different types of underwater targets effectively and generate feature vectors with high degree of division.It also reflects the high application value and practical significance of motion feature in underwater target recognition.(3)From the comprehensive analysis of feature similarity,classification recognition rate,feature dimension and other aspects,it is proved that 3D Zernike velocity moments has better performance in the recognition of fish and ROV binary classification,and the recognition results are better than the traditional methods.