Synthetic Aperture Sonar Image Mosaic Technology Based On Feature Extraction

Synthetic Aperture Sonar technology,or SAS,is a type of side-scan Sonar.The characteristic of SAS is that the resolution in the azimuth direction is independent of the target distance.The realization principle of SAS is to simulate a virtual large aperture by using the coherence of sound waves in water after the base array uniformly moves upwards in the direction,and finally make the resolution of the azimuth direction independent of the target distance.SAS for submarine topography detection as well as the development of Marine safety has a vital role.However,the images generated by sonar system are generally banded,which is not conducive to direct observation.In this paper,the SAS image Mosaic technology and the corresponding algorithms are deeply studied,and combined with the characteristics of SAS technology,some algorithms are optimized.The main content of this paper is as follows:The realization principle of SAS is discussed,and the mathematics proves that SAS can achieve high resolution in azimuth direction.Several imaging algorithms of SAS are analyzed,including point-by-point imaging algorithm and line-by-line imaging algorithm.Although the point-by-point imaging algorithm is accurate and easy to understand,its complexity is too high,which will affect the imaging speed of SAS system.Line-by-line imaging algorithms mainly include Range-Doppler algorithm,Chirp-scaling algorithm and wave-number domain algorithm.Among them,The RangeDoppler algorithm utilizes Taylor expansion for approximation,while the Chirp-scaling algorithm and wave-number domain algorithm adopt Stolt transform.The difference is that the Chirp-scaling algorithm is approximate while the wave-number domain algorithm is accurate calculation.This leads to the chirp-scaling algorithm which may seriously affect the imaging results when the observation angle is too large.In terms of Feature extraction,the performance of scale-invariant Feature Transform(SIFT)algorithm,Speed up Robust Features(SURF)algorithm,Oriented Fast and Rotated Brief(ORB)algorithm and deep learning algorithm based on superpoint structure,as well as their advantages and disadvantages in sonar image processing are analyzed.The above four algorithms possess rotation invariance and scale invariance.Compared to SIFT algorithm and SURF algorithm,the ORB algorithm and the deep learning algorithm of feature extraction rate is faster.However,since ORB algorithms extract feature points with a high degree of distinction first,this can lead to a cluster of feature points.In this paper,in order to improve the ORB algorithm,a strategy of using quadtree to uniformly extract feature points was proposed.Finally,the image mosaic was completed without affecting the rate of algorithm.In the aspect of feature point matching,the K-nearest neighbor algorithm based on balanced KD tree,KNN algorithm for short,and the weighted optimized KNN algorithm are proposed.At the same time,the RANSAC algorithm was used to optimize the matching pairs,and finally the homography matrix between matched images was generated.A uniform motion model is proposed to optimize the matching of feature points according to the low local differentiation of sonar image and the possibility of image shadow.Then,the selection strategy of image fusion algorithm on sonar image is discussed.At last,the work of this paper is summarized.