Semantic Regularization for Blind image Deconvolution in Medical Images with extended sparseness

Blind Image deconvolution is one of the key areas of research and is a challenging task in data restoration. Many deconvolution techniques have been proposed and their evaluation as a successful counterpart in the area of medical imaging is still void. Unknown cause for degradation and amount of degradation is of main concern and is solved by regularization of image features using semantic nets with extended sparseness in our proposed model. A degraded image is convolved using a kernel matrix based on Latent semantic analysis and the resultant co-occurrence matrix is regularized using Tikhonov regulariser. In order to reduce the reconstruction error a non-negative matrix factorization approach is used resulting in a better quality image post regularization. Non negative matrix factorization aids to retrieve high level semantic features. Based on this assumption matrix factorization is used to model the blur model (Point spread function (PSF) is known, kernel matrix is factorized to find the semantic features) and the extracted features are related with the existing features for further extraction. Improved semantic networks with extended sparseness are analyzed for two most popular convolutional semantic techniques namely Latent semantic analysis and probabilistic latent semantic analysis. Feasibility of extended sparseness for blind image deconvolution over the two methods is tested on real time ultrasound and dermatology datasets and the performance of the same is justified using performance metrics PSNR, SSIM, BIF and VSNR. Our proposed method results in images with better PSNR and structural similarity index values compared to original images.