Zhenmin Zhu

Image reconstruction algorithm from compressed sensing measurements by dictionary learning

Abstract It is a challenge task to reconstruct images from compressed sensing measurement due to its implicit ill-posed property. In this paper, we propose an image reconstruction algorithm for compressed sensing image application based on the adaptive dictionary, which is learned from the reconstructed image itself. The sparsity level is enhanced since the sparse coding of overlapping image patches emphasizes the local image features; accordingly the quality of the reconstructed image is also improved. In addition, Batch-OMP algorithm, linearization technique and dynamic updating sparse coding algorithm are used to reduce the computational complexity of our proposed algorithm. Numerical experiments are conducted on several test images with a variety of sampling ratios. The results demonstrate that our proposed algorithm can efficiently reconstruct images from compressed sensing measurements and achieve more than 3xa0dB gain averagely over the current leading CS reconstruction algorithm.

Classification-Based Adaptive Compression Method for Computer Screen Image

In this paper, a classification-based adaptive compression method for computer screen image is presented. This method firstly divides the computer Screen Image into 16×16 non-overlapping blocks, and then every block is classified into three types: text/graphic, pictorial and hybrid blocks based on the characteristics of histogram distribution and the number of colors. For complex text/graphic block, k-Means clustering method is used to reduce the number of colors to improve compression performance, finally the text/graphic block is coded by our proposed lossless coding method, hybrid block is coded by hybrid coding method and the pictorial block is coded by H.264-like intra coding method. Experiment results show that our proposed block classification method exactly distinguishes three block types, color clustering method can effectively reduce the number of colors for complex text/graphic. The compression performance and subjective image quality of our proposed method can outperform JPEG and JP2k.

Motion estimation for video coding based on sparse representation

This paper describes a motion estimation algorithm based on sparse representation, which can be applied in video coding to reduce the temporal redundancy. The sparse coefficients are firstly calculated in support region by orthogonal matching pursuit (OMP) algorithm using the reference blocks as dictionary elements, and then these optimal sparse coefficients are utilized to predict the current block. To get the same prediction in decoder, the number of iterations in OMP is transmitted to decoder as side information. Simulation results show that gain up to 2.87dB in terms of the PSNR when compared with traditional translational motion estimation model.

Improved Compressed Sensing Image Recovery Algorithm by Intra Prediction

In recent years Compressed Sensing (CS) has drawn quite an amount of attention as novel digital signal sampling theory when the signal is sparse in some domain. However, signal recovery from compressed measurement data has always been challenging due to its implicit ill-posed nature. In this paper we will propose an improved CS image recovery algorithm by intra prediction method based on block-based CS image framework. The current block is firstly predicted by its neighbor pixels, and then its prediction residual is recovered. The performance of our proposed CS image recovery algorithm is superior to the traditional CS recovery algorithm because the sparsity level of prediction residual is higher than its original image block. Experimental results demonstrate that the proposed algorithm outperforms the traditional CS image recovery algorithm and achieve by far 2dB gain in PSNR.

Image compressed sensing reconstruction with 3D transform domain collaborative filtering

Compressed Sensing (CS) has drawn quite an amount of attention as novel digital signal sampling theory in recent years when the signal is sparse in some domain. However, signal reconstruction from undersampled data has always been challenging due to its implicit ill-posed nature. This paper proposes an image compressed sensing reconstruction algorithm for image CS application, which consists of iteratively collaborative filtering of non local similar image patches in 3D transform domain and solving the least squares problems. In addition, the linearization technique is exploited to reduce the computation complexity. The results of various experiments on natural images and MRI images consistently demonstrate that the proposed algorithm can efficiently reconstruct images and gain more 2dB as compared to the current leading CS image reconstruction algorithm.

Compressed Sensing Image Reconstruction Algorithm by Dictionary Learning

It is a challenging task to reconstruct images from compressed sensing measurement due to its implicit ill-posed property. In this paper, we propose an image reconstruction algorithm for compressed sensing image application based on the adaptive dictionary, which is learned from the reconstructed image itself. The sparsity level is enhanced since the sparse coding of overlapping image patches takes into account the local image features, and accordingly the quality of the reconstructed image is improved. In addition, linearization technique is also exploited to remove the computation of matrix inversion. Numerical experiments are conducted on several test images with a variety of sampling ratios. The results demonstrate that our proposed algorithm can efficiently reconstruct images from compressed sensing measurements and achieve more than 3dB gain averagely over the current state-of-art compressed sensing reconstruction algorithms.

Motion Estimation for Video Coding Based on Subspace Pursuit

In this paper, a motion estimation algorithm is proposed based on subspace pursuit, which can be used in video coding to reduce the temporal redundancy. The main idea of our proposed algorithm is that the process of motion estimation will be imagined as sparse representation, that is, the predicted block is grouped as an observation vector and the corresponding reference blocks in previous reconstructed frame are used to construct a sparse dictionary, then the sparse coefficients are calculated by subspace pursuit algorithm. In order to reduce the transmission of sparse coefficients, the idea of template matching is adopted. Moreover, our proposed method can be combined with traditional motion estimation algorithm to further enhance the inter prediction accuracy. Simulation results show that our proposed method can outperform the decoder motion vector derivation method in term of the peak signal to noise ratio (PSNR).