Research On Sparse Representation And Its Application In Underwater Vision Navigation Data

The accurate navigation is the precondition of fulfilling tasks for underwater vehicle.However,the position of the vehicle needs effictive information to revise after long time navigation with methods such as inertial navigation,dead reckoning and simultaneous localization and mapping.It is great importance of synthetic aperture sonar(SAS)image denoising and recognition since underwater vision navigation based on high resolution SAS images provides precise location,which assists the vehicle navigation accurately.For the differences of imaging encironment and way,sonar image processing with the optics methods usually has a bad effect.As is known to all,mammals represent image in visual cortex just using a handful of neurons within the field of vision.Similarly,sparse representation theory received extensive attention for its capability of representing signal concisely.In recent years,sparse representation has made certain progress in application of the inverse problem such as image restoration,blind signal separation,medical imaging.In this paper,the dictionary learning and sparse coding algorithms via sparse representation are first studied.Then,SAS image denoising and identification via sparse representation are researched with its characteristics.The main research contents in this paper are as follows,The Intelligent Block K-SVD algorithm is proposed to get the optimal dictionary size for balancing the representation error and training time.For the block structure in image is usually unknown in practical application,sparse clustering algorithm is carried out to solve by dividing the similar coefficients of the sparse representation in a group.Then,the initial dictionary size is setting based on the block sparsity of the image,and atoms are added for dictionary expansion.The dictionary trained by the proposed algorithm has a smaller size than that by other methods with the same representation error.The intelligent Block K-SVD algorithm also has a shorter training time than most current dictionary learning algorithms.In many practical applications,the sparsity of the signal or image is often unknown,and cannot be estimated accurately.For traditional sparse coding algorithms in synthesis model executes with sparsity as a priori information,stagewise regularized orthogonal matching pursuit algorithm is proposed in this paper.First,the signal sparsity is estimated and the candidate set is established based on the stagewise principles.Then the candidate set is regularized to recover signal and update the residual.The proposed algorithm has a better performance on reconstruction speed and precision and is possible to be used in application.The dictionary size of large scale and the way of atoms updating are two important factors that impact the training speed of analysis K-SVD algorithm for operator learning.Therefore,Parallel Atom Updating algorithm is proposed to accelerate by replacing reconstructing signal with establishing cosupport.The atoms updating is also simplified with submatrix multiplication instead of SVD decomposition to short the training time.The proposed algorithm is improved obviously in learning speed and robustness to noise.The current analysis cosparse coding algorithm needs cosparsity as a priori information when recovering signal.However,the signal cosparsity can`t be estimated accurately.Therefore,the correlation between the sparse coding algorithms of analysis model and synthesis model is first analyzed,and the cosparsity-based stagewise matching pursuit algorithm is proposed as an improved algorithm.An effective mechanism for estimating the signal cosparsity is carried out,then approximating the real cosparsity with a large step size until meeting the stopping condition.The fine-tuning of the cosparsity is executed in the backward process to recover signal.It is verified both in theory and experiment that the proposed algorithm is stable and robust for the recovery of noiseless and noisy signal.For the characteristic of SAS images is different from optical images,SAS image denoising with the common methods is hard to obtain a good result.Hence,denoising method for SAS image based on sparse representation is proposed here.The additive noise and multiplicative noise in SAS image is removed in two stages.The denoising dictionary is trained via the noisy image,and then the noise can be moved since the image can be represented sparsely but the noise can't,so the SAS images get a precious reconstruction.Image recognition based on sparse representation have advantages in large number of training samples,however,the algorithm cannot be directly used in sonar image recognition for SAS image is difficult to require.In this paper,image recognition algorithm based on sparse representation is improved first,and then partitioning the sample data is proposed to increase the number of samples.Different from the traditional way of composing the training sample matrix,the sub-dictionary for the same class is trained according to all kinds of samples and then merged to be the sample dictionary.The improved algorithm can complete the image recognition with high recognition rate with SAS image in less quantity.