Research On Data Gathering,Restoration And Compression Based On Sparse Representation

With the fast development of information acquisition technology and network technology, digital signal grows exponentially with sensor data and image/video as its typical representatives, which poses great challenge and pressure to the applications such as data gathering, image restoration and image/video compression. Sparse representation theory provides new avenues for efficient representation and accurate reconstruction of digital signals, however, aiming at the applications such as data gathering, image restoration and image/video compression, how to fully utilize the property of signal itself and mine the prior knowledge based on sparse representation, to establish efficient signal representation model and robust reconstruction model, it is still an open problem, and related research is of great theorical and realistic significance.Based on the sparse representation theory, this dissertation investigates the applications of data gathering, image restoration and image/video compression, and proposes a series of models and schemes to improve their performance. These schemes can be classified into the following three categories:First, in the aspect of wireless sensor networks(WSN) data gathering, we propose data gathering scheme based on adaptive compressive sensing(CS), and establish adaptive reconstruction model based on auto-regressive(AR) model to achieve efficient data gathering. To solve the problem that current CS based data gathering schemes are not efficient enough to deal with complex and changing sensor data, we propose AR model based adaptive reconstruction scheme by exploiting the high correlation between neighboring sensor nodes in WSNs. Furthermore, to solve the problem of prevalent anomaly in WSNs, we propose an anomaly recovery and detection scheme based on joint sparsity of transform domain and spacial domain. To deal with the problem that existing CS based data gathering schemes use fixed number of measurements, we propose an adaptive measuring scheme based on recovery error estimation by utilizing the fact that random CS measurements are of equal importance and the progressive property of CS recovery. Experimental results on real WSN datasets show our scheme significantly improves the quality of recovered sensor data.Second, in the aspect of RGB color image and depth image(RGB-D) enhancement, we propose group sparsity recovery model of multi-channel imgae and orthogonal sparse dicionary learning method for multi-channel images. By analyzing the high correlation among different channels of RGB-D image, we found the sparse coefficients of different channels share nearly the same support. We propose multi-channel image group sparsity recovery model according to such property. To solve the problem of low computational efficiency in sparse coding due to current overcomplete sparse dictionary, we propose multi-channel orthogonal sparse dictionary learning scheme for RGB-D image. Thus sparse coding problem has closed-form solution given orthogonal dictionary and we prove it theorically. The computational complexity is thus dramatically reduced by replacing traditional iterative solutions with closed-form solutions in sparse coding. Experimental results show the proposed scheme significantly improves both the subjective and objective recovery results for RGB-D images.In the aspect of image interpolation, we propose a local and non-local joint low-rank recovery model for images. Based on the low-rank recovery theory, by exploiting the local correlation and non-local self-similarity in images simutltaneously, we propose a local and non-local joint low-rank recovery model. The image interpolation problem is then reformulated as a low-rank matrix recovery problem. We also provide the corresponding efficient solution algorithm based on Split Bregman algorithm. Experimental results show the proposed scheme efficiently improves both the subjective and objective quality of interpolated images.Third, in the aspect of intra prediction in video coding, we propose an intra prediction scheme based on low-rank matrix completion. To deal with the problem that standard H.264/AVC cannot predict complex texture well, we establish the low-rank representation model for video frames by exploiting the global self-similarity in frames. The similar image patches are searched by template matching and rearranged into an approximately low-rank matrix. The intra prediction problem is thus reformulated as a low-rank matrix completion problem. We define directional templates for template matching to further capture the texture property of patches precisely. Experimental results show the proposed scheme outperforms the standard H.264/AVC in both the subjective and objective result.In the aspect of image compression, we propose an image compression scheme based on the sparse representation of primal sketch. To solve the problem of high frequency information loss in current image compression methods, from the perspective of computer vision, we propose to restore the lost high frequency information by exploiting the primal sketch prior information learnt from training image sets. The efficient and compact representation of trained primal sketch from training image set is obtained by utilizing orthogonal sparse dictionary learning. By exploiting the structure similarity between high quality and low quality primal sketch pair, together with a learnt primal sketch orthogonal sparse dictionary, the lost high frequency information is restored by sparse reconstruction. Experimental results show the proposed scheme outperforms JPEG 2000 at low bit-rate in both subjective and objective quality, and better subjective quality can be achieved especially at edge areas in images.