Studies On Nonlinear Image Compression Algorithm Based On Dictionary Learning

Image compression is one of the important research contents in the field of image processing.Effectively compress images containing large amounts of data has important research significance.Therefore,aiming at the shortcomings of dictionary learning in the field of image compression,this thesis conducts the following research works:1.Considering the limitations of concentrated dictionary learning algorithm in denoising ability and complexity,the image compression algorithm based on denoising deep extreme learning machines based on autoencoder and approximate K singular value decomposition is proposed.Aiming at the problem of inadequate denoising ability in the concentrated dictionary learning algorithm,the proposed algorithm uses denoising deep extreme learning machines based on autoencoder to obtain high-level feature representation of data,and obtains denoising dictionary by approximate K singular value decomposition,thereby,the denoising ability of the algorithm is improved.Furthermore,denoising deep extreme learning machines based on autoencoder and approximate K singular value decomposition are used to further reduce the complexity of traditional concentrated dictionary learning algorithm.Experiments show that the improved algorithm proposed in this thesis is superior to the concentrated dictionary learning algorithm in terms of denoising ability and complexity.2.Aiming at the problem that vectorization affects data structure in tensor signal processing,a tensor image compression algorithm based on sparse representation is proposed.Dimensional dictionary matrix and sparse coefficient tensor of input tensor signal are obtained by sparse decomposition.Core coefficient tensor and factor matrix are got by Tucker decomposition of tensor signal.On this basis,a special tensor sparse representation is established by approximating the relationship between the sparse coefficient tensor and the core coefficient tensor.The concentrated energy dictionary learning algorithm is introduced to reduce the dimension of the dictionary in the sparse representation to achieve tensor compression.Experiments show that the proposed algorithm has significant advantages in preserving the original data information and denoising ability compared with other comparative algorithms in hyperspectral image data compression.