Research On Adaptive Block Compressed Sensing Based On Image Feature

The whole 2-D natural images are directly sampled when traditional methods of compressed sensing are adopted.The main drawback of the traditional approaches is the computation and storage of the observation matrices since large-scaled observation matrices are required to cover the whole images.To bridge this gap,the block compressed sensing which has the whold images quickly sampled and reconstructed and guarantees real-time performance,was introducted.However,when the traditional block compressed sensing is used,the constant sampling rate for all sub-blocks of a image leads to unbalanced sampling effort since the information varies from block to block.For the problem at hand,several improvements are made for the traditional block compressed sensing in this thesis:(1)Firstly,two adaptive sampling strategies are reviewed based on gray entropy and edge information,respectively,according to the knowledge that the amount of information varies from block to block.Those two strategies have each sub-block set up with a proper sampling rate according to the feature of pre-estimated images that obtained by pre-sampling.Those two strategies result in two block compressed sensing algorithms with adaptive sampling rate in the spatial domain.With the aid of the above two algorithms,a new block compressed sensing algorithm with adaptive sampling rate is proposed based on total variation.More precisely,the sampling rate for each sub-block is assigned according to total variation.Numerical results show that improved reconstructed image quality and visual effect is achieved by the block compressed sensing algorithms with adaptive sampling rate based on the image features with slightly more computational cost when compared with the traditional block compressed sensing algorithm,and the proposed block compressed sensing algorithm with adaptive sampling rate based on total variation outperforms its counterparts.(2)The multiscale block compressed sensing has the performance of the traditional block compressed sensing improved by transforming the image into the multiscale wavelet domain and assigning proper sampling rate adaptively for each block according to different scales of wavelet coefficients.However limitation of multiscale block compressed sensing exists since the same sampling rate is applied for the blocks with the same scale of wavelet coefficients and the low frequency coefficients that contain a large amount of prior information are not fully used.Motivated by this shortcoming,the multiscale block compressed sensing with adaptive sampling rate based on gray entropy and edge information in the wavelet domain are studied with all the low frequency coefficients reserved.Based on those reserved coefficients,pre-estimated images are obtained and then proper sampling rates are adaptively assigned among those sub-blocks with the same scale of wavelet coefficients with the aid of the pre-estimated images.When the total variation is introduced,the multiscale block compressed sensing algorithm with adaptive sampling rate based on total variation is proposed in this thesis.Numerical results show that improvement is achieved by the the multiscale block compressed sensing with adaptive sampling rate based on the image features in term of reconstruction quality when compared with the traditional multiscale block compressed sensing algorithm.Besides,the proposed multiscale block compressed sensing algorithm with adaptive sampling rate based on total variation outperforms its counterparts.