JOINT SPARSE CONVOLUTIONAL CODING FOR IMAGE SUPER-RESOLUTION RESTORATION

The globality of the image is a strong constraint in the process of image super-resolution reconstruction. Convolutional sparse coding uses the global features of image in super-resolution reconstruction, and decomposes image wholly into convolution sum of filters and feature map instead of encoding the image patches. However, sparsity guided by the 1(l) norm cannot represent high-order structured information. In this paper, the joint sparse convolution coding (JSCC) proposed to further extract high-order structured information, leading to a compact dictionary. Our model involves the three sets of parameters to learn, the decomposition filters, mapping function and reconstruction filters. The model utilizes structured sparse regularization term l(2,1) regularization to constrain feature maps instead of single l(1) norm. Meantime, extract edge prior information for registration with high frequency components. Extensive experiments demonstrate JSCC model can achieve competitive PSNR results, while reconstruction image illustrates better texture preservation performance and edge information. l(2,1) regularization term can also avoid the influence of noise, reflecting the superior robust performance.