Denoising Via Sparse Representation And Structure Clustering For WMSN Image

WMSN is mainly composed of mass multimedia sensor nodes with computation, storageand wireless communication，which possesses the advantages of sensing rich information andstrong extensibility. WMSN is usually deployed in an unattended environment monitoringtask independently and widely used in many fields. In addition, higher requirements are putforward on the accuracy and effectiveness of monitoring to complete the application invarious video monitoring tasks. However, due to the complicated WMSN monitoringenvironment, the video quality is susceptible to external factors such as bad weather,illumination change. Then the image is affected by noise pollution so seriously that the visualeffect is fuzzy. This is so easy to cause unpredictable accidents.In order to ensure the accuracy and effectiveness of the video monitoring, high qualityimage does great significance on the subsequent image processing and monitoring. After thefurther analysis，there is a large amount of data information in the WMSN video imageacquired, the target image is very similar and the continuous frame image of the same goal isstrong correlation and has much redundant information. The study found that image sparserepresentation, which can describe essential feature of the image to fully implement efficientrepresentation, is widely used in image and video denoising. In addition, the non-localdenoising method makes full use of the image self-similarity and redundant features, whichcan fully keep the details of the image texture structure, and the denoising effect is great.Combining with the WMSN video image and making thorough analysis of sparserepresentation and nonlocal theory, it is very necessary to study suitable denoising algorithmfor WMSN video image under complex scenarios.On the basis of the previous study, combining with sparse representation and nonlocaltheory, this paper deeply studied WMSN video image denoising algorithm, two kinds ofimage denoising algorithm which are suitable for WMSN scene are put forward. The first isan image denoising algorithm based on structural clustering and centralized regularization. Inthe process of this algorithm implementation, the degraded original WMSN images arerespectively structurally clustered by K-means clustering based on similar geometric structurefirstly. Then, on the basis of the thought of double sparsity, the dictionary is trained throughsparse K-SVD in order to enhance the structural dictionary. At last, the centralizedregularization is performed to optimize the sparse coding coefficients. These steps leads tomore structure information reserved and be better reconstructed. Experimental results showedthat in low SNR condition this algorithm can ensure that large edge, texture and structureinformation are reserved. Moreover, structure clustering and multiple dictionary parallel processing can speed up the algorithm running speed and reduce the WMSN network energyconsumption. The other is a WMSN image denosing method combining a nonlocalMarkov-chain Monte Carlo sampling and adaptive threshold low-rank approximation. In theprocess of this algorithm implementation, firstly, the cluster of similar patches is searched byMarkov Chain Monte Carlo (MCMC) sampling. Afterwards, the singular values are designedwith different thresholds according to the image prior information. Then, the singular valuedecomposition is conducted; At last, image reconstruction is done with low rankapproximation algorithm. Experimental results clearly show that the proposed outperformsother denoising algorithms in terms of quantitative measure and visual perception quality. Thealgorithm improves the image edge, texture and structure information retention. In addition,this method solves the limitations of the dictionary learning in dealing with high-dimensionaldata and reducing the computation complexity. Consequently, this algorithm is very suitablefor energy constrained WMSN video image denoising.