Study On Image Recognition Technique Of High Resolution Imaging Sonar

This paper carries out with the project named "underwater target acoustical detection and recognition" in the ten five-year national defense study in advance, which is a part of the military intelligent underwater vehicle. With the view to the engineering problems, the paper makes a deep study on the structure and frame of the acoustical vision system, and images pre-processing, feature detection and recognition of side-scan sonar and forward-looking sonar. The major contents are as follows in general:(1) The paper gives a survey of the internal and external current status of research and progress trends on underwater intelligent vehicle, underwater target recognition, and the major equipments or methods. It also introduces the frame of current image recognition system and primary technology particularly.(2) We also survey on performance and characteristic of current international advanced high resolution imaging sonar——high resolution side-scan sonar, acoustic lens sonar and 3D imaging sonar. Then we designs a new structure of acoustical vision system for the major mission of military underwater intelligent vehicle. With it the vehicle has the ability to detect forward and downward simultaneously and recognize targets "coarsely" faraway and "finely" close. So it can detect and recognize targets more effectively.(3) The noise disturbance of sonar imaging is analyzed. The paper makes a deep study on relative theories of pulse coupled neural net(PCNN) and proposed a algorithms of incorporated morphology and median filter based on simplified PCNN modol. This method doesn' t destroy the image edge and wipe off the gaussian and pulse noise simultaneously. With the inherent parallel capability of PCNN, the method is more suited for real-time processing.(4) We introduce the major application of side-scan sonar in vehicles and survey on the new development of seafloor sediment research with side-scan sonar images. We propose to combinate the phase spectrum and gray level co-occurrence matrix to denote the image texture, and obtain the feature vector by particular component decomposition(PCA). With the incorporation of statistic and structural information, the recognition ability is much higher.(5) For the reason of different images with different space distribution of gray levels, we propose a texture denoting method based on simplified PCNN model which obtain a series of images corresponding to different gray levels. Then transforming image series into 1D variance series to form feature vector with operating variance of each image. The feature is rotation invariant and provides high recognition rate for sonar images and natural texture images classification.(6) To deal with the problem of seabed reverberation of side-scan sonar images in target detection, we propose to estimate the targetin images by the magnitude of higher-order spectrum----bispectrum.This method can restrain the background noise in images and enhanced the target information, so it can detect target in different side-scan sonar images universally.(7) We segment the side-scan sonar images to detect targets by combinating autonomous grey level adjustment and Otsu' s method. It can auto-select threshold and proves better segmentation result represented by region uniformity and contras.(8) The paper analyzes the characteristic of high resolution forward-looking sonar, and propose a morphological edge-detection method which isn' t sensitive to noise and can eliminate gray burst caused by uneven echo signals. As a result, it can get integral contours and get rid of isolated points.(9) To deal with the problem of image rotation, translation and scale variance caused by vehicle motion, we make a deep research on the vision invariant theories of image recognition and propose a invariant feature detection based on discrete Radon transform of image edges after scale normalization. Radon transform can project 2D edge into 1D space, then we construct moment invariant in the space and use singular value decomposition (SVD) of different order moment invariant matrix to detect target feature. This method is robust and have higher recognition rate, and because only use the edge information, so operating is faster.(10) At the end of the paper, we give a survey of 3D target recognition technology, the keypoint is discriptions of 3D target and recognition methods.Then we illuminate process of recogniton of 3D imaging sonar based on the principal of 3D sonar imaging, using the methods of 3D reconstruction and recognition in industry and medicine. Finally we introduce the key technologies in recognition and analyzing the main difficuties in real tasks.