Motion Estimation And Denoising Fusion Based On Sonar Image

Sonar is a critical instrument in underwater detection using AUV.With the development of high-resolution forward looking sonar,underwater detection techniques based on sonar images draw people's attention more and more.However,free floating of the vehicle will result in distortion in gray level and geometrical structure.Texture features in sonar images are weak in contrast to optical images.This paper focuses on two typical application problems in underwater detection.A sonar image motion model describing the relationship between 2-dimension sonar image and 3-dimension vehicle motion is built based on sonar imaging principle.Image matching methods taking advantages of characteristic point distribution probability model are proposed for seabed images and motion parameters are estimated by serial images.Directional wavelet is employed in image analysis for underwater large targets.Two image fusion methods are built based on directional filter and morphology.Sonar image motion model is a model in which 3-dimensional differential motion of the vehicle is used to describe the 2-dimensional characteristic points.Foundation of the model is sonar imaging principles,including 2-dimensional target motion models approximately describing 3-dimensional vehicle motion,3-dimensional motion models describing 6 degrees of freedom vehicle motion and 3-dimensional beam image motion models based on original sonar data.Meanwhile,elevation angle estimation method is studied and motion parameters are estimated by shadow movement in images.Sonar image matching and motion parameters estimation methods are studied with the help of sonar image motion models for seabed images.A framework for image matching is built based on the characteristics of sonar images.Firstly,inconsistency in gray level distribution is compensated and feature extraction methods for stable characteristics and interested zones using gradient and gray level threshold are proposed.Secondly,stable probabilistic model are used in the study of regional Gauss probability for feature points distribution description which can describe the discrete points in sectional continuous models.Finally,matching framework derivation for motion parameters and feature points are exhibited.By constructing the optimization model of regional probability,image transformational matrix or motion parameters are calculated and a fast method is proposed.To study the accuracy of this estimation,pool experiment is carried out and many serial images are matched for accumulative error analysis.An embedded sonar detection platform is designed for ROV which will be the basement for next algorithms.Directional multi-scale image fusion methods are studied for large target surface and structure details.Firstly,some directional multi-scale wavelets are discussed for the application in sonar image fusion,including ridgelet transform,curvelet transform and contourlet transform.Secondly,a sonar image fusion method(NSCTMM)is proposed using NSCT and morphological decision.Finally,morphological wavelet and directional filter are combined in order to give out the MWTD.A detection experiment for the surface of convex steel plate is taken.Sonar images of underwater oil platform and steel plate are analyzed and the results show that NSCTMM performs well in information-preserving,target connectivity and gray level richness.MWTD is weaker in detail fusion but it also can extract enough information and work well in noise reduction.All these research are of great significance for target detection and recognition in the future.