Single Image Super-Resolution via Convolutional Sparse Coding with One Group of Filters

Using the characteristic that image patches can be represented sparsely on a selected over-complete dictionary, the sparse coding based super-resolution (ScSR) method pioneers a new avenue for the SR technology. However, ScSR will result in obvious block-effect because of the block preprocessing of the image. For this purpose, Zeiler et al. proposed convolutional sparse coding based SR (CSC-SR) method to solve this problem. Because it needs to learn two groups of filters and mapping function between HR and LR feature maps, CSC-SR usually puts forward high demand to storage and relatively large the calculation expense. In this paper, we propose a modified CSC-SR method to single image super-resolution. The proposed method only needs to train one group of filers, and explores directly the mapping function by least squares algorithm, which can effectively reduce the workload of model training. Experimental results for different test images show that our model has achieved effective improvement over the previous methods.