Denoising Via Sparse Representation And Dictionary Learning For WMSN Video Image

As an important technology of Internet of Things to active perceiving the physical world,WMSN technology has been widely used in terms of the unique advantage of perceivingmultimedia information. However, the complexity and high noise level of the monitoringenvironment lead to the result that video images often contain severe noise. Therefore, videoimage denoising becomes the key to ensure the validity and reliability of the WMSN videomonitoring. Recently sparse denoising has drawn an increasing number of research attention,which can portray essential characteristics and distinguish the useful information from thenoisy image. And the sparse representation based on over-complete dictionary, largelydetermining whether the image characteristics can be represented or not, has been becomingthe research focus in this field.According to the analysis on the characteristics of the WMSN video image and combingthe sparse decomposition, the sparse denoising algorithm based on the K-SVD algorithm andthe residual ratio for WMSN video image in low SNR is proposed. Firstly the features ofWMSN video images are analyzed. And then, according to the above analysis and the sparsedecomposition theory, the redundant DCT dictionary is adaptively trained by the K-SVDalgorithm to reflect the image structure characteristics. At last, under the situation of low SNR,the video image is reconstructed through the Batch-OMP algorithm using the residual ratio asthe iteration termination. Experimental results showed that the proposed algorithm couldeffectively filter out the strong noise, suitable for the WMSN video image denoising.Based on the previous study and for the purpose of fully exploiting nonlocalself-similarity of WMSN images, the sparse denoising algorithm via clustering-based sparserepresentation is proposed. Firstly, WMSN images are respectively clustered based on thepixel intensity of Regions of Interest (ROIs), which are determined in terms of Bayesiantheorem. Secondly, in the light of non-local self-similarity regularizer provided by theROI-based WMSN images clustering, Clustering-based Sparse Representation (CSR) builds anew sparse denoising model exploiting both sparsity and nonlocal self-similarity to improvethe quality of reconstructed images. At last, a surrogate-function based iterative shrinkagesolution has been developed to solve the double-header l1-optimization problem.Experimental results showed that the performance of our approach to denoising is competitive,qualitative as well as quantitative, suitable for the WMSN video image denoising.