Research On Underwater Image Segmentation And Typical Target Extraction And Recognition

Underwater images have been widely used in the fields of military affairs and civil matters.Thus the corresponding technologies such as underwater image processing and target recognition have become a hot area of research.The process of underwater target recognition usually contains four parts:underwater image preprocessing,image segmentation,feature extraction and target recognition.The images,which are acquired by underwater optical imaging system,are degraded to some extent due to the absorption and scattering of suspended particles.Underwater images share certain characteristics of low contrast,noisy and intensity inhomogeneous,which bring challenges to post-processing such as image segmentation and object recognition.On the other hand,the targets under water are usually man made objects,for example,torpedoes,submarine and submarine pipelines.Those targets commonly have weak texture features,but strong shape features.This dissertation focuses on active contour models for images segmentation and support vector machine.We simulate two groups of models through Solidworks:one of which contains four typical underwater objects which are ellipsoid,cylinder,sphere and cube,the other of which consists of submarine,torpedo,mine and cube.The whole process consists of image pre-processing,image segmentation,feature extraction and object recognition.The main works in this dissertation are outlined as follows:1.We firstly introduce the vision system model and analysis the characteristics of underwater images.Then,on the basis of comparing the related algorithms,we choose the appropriate approaches for underwater images enhancement and denoising.We perform the background correction model and adaptive histogram equalization respectively to deal with intensity inhomogeneity.In addition,the gray-scale transform is used to improve contrast.Aimed at the problems caused by noise,we implement the adaptive dictionary learning algorithm K-SVD not only to achieve the denoised images,but also to retain the details simultaneously.2.The research core of the dissertation lies in the segmentation methods which are based on active contour model and level set theory.Firstly,we propose an online active contour model for specific object(s)segmentation,whose energy functional is only constructed by the statistics inside the initial contours.Thus our model has the ability of detecting the object(s)assigned by the initial contour,as well as those which share the same intensities.Secondly,given that classical CV model is based on the difference of the pixel and region average intensities,it would become invalid when it is polluted by noise.Accordingly,we describe an improved region-based CV model in a variational level set formulation,whose basic idea is to exploit the difference of local statistical information and global image information inside and outside of the contour.In addition,the proposed model is extended to multi-phase one.Furthermore,the Gaussian filtering processing is adopted to regularize the level set functional and reduce the affect of the noise to a certain degree.Thirdly,we combine the global and local statistical information in a new way to define a fitting image.With level set representation,the energy functional is based on the difference between the fitting image and original image.Thus the hybrid model has capacity of handling with intensity inhomogeneity and robustness to initialization.3.Consider that the typical underwater targets have typical shape features,we construct combined moment invariants which are in possession of translation,rotation and proportion invariance.The combined invariants are composed of NMI,improved Hu relative invariants and affine invariants.Then we apply the principal component analysis(PCA)and independent component analysis(ICA)algorithms to reduce the redundancy of invariants,and thus to realize feature dimension reduction and feature optimization.Finally,by using Solidworks software,we simulate two groups of objects to get 3-D diagrams.Afterwards,a large number of silhouette images through different perspectives and direction could be obtained.The simulation results have demonstrated the excellent performance of invariants and the feature selection ability of PCA and ICA.4.Target recognition is the critical step of the whole vision recognition system.Up to now,there have been kinds of classifiers,among which support vector machine is one of the most effective ones.In this dissertation,we propose an improved projection twin support vector machine,whose model is to seek two projections directions,one for each class,by solving a single quadratic programming problem.The projected samples of each class are well separated from those of the other class in its respective subspace.In order to further boost performance,a recursive algorithm is generated more than one projection axis for each class.Besides,we adopt an efficient clipping dual coordinate descent(DCD)algorithm to solve the dual problem of our model.Our linear model can also be extended to a nonlinear case with the kernel trick.Experimental results on several public dataset show the significant advantages of our model in terms of both classification accuracy and computational complexity over the state-of-art methods.Furthermore,the application to the simulated underwater objects data has also shown the efficiency of the proposed approach.5.We conduct experiment in terms of the underwater typical target recognition system.Firstly,the hardware structure is constructed for typical targets recognition system.Then,we take Lab VIEW and IMAQ VISON as the software platform.The experiments results have demonstrated the ability of the target recognition system.The effect on the pictures taken by the vidicon has validated the performance of image enhancement algorithm,the denoising ability of K-SVD,three kinds of active contour models,the clustering capability of the combined invariants and the higher accuracy of improved projection twin support vector machine.