Learning based image super-resolution using contour stencils as edge model

In this paper, we present a new learning based technique to super-resolve a low resolution image using “contour stencils” and a database of low resolution (LR) images and corresponding high resolution (HR) versions. The curves in the test image are modeled by fitting the stencils that give smallest total variation along the curves. We search the best matching low resolution contour stencils from the LR training images in the database and obtain high resolution curves from corresponding HR training image. We reconstruct the super-resolved image by combining the learned high resolution information and the observed low resolution information. The experiments are performed on real world images. The results of the proposed method are compared with that of the standard interpolation techniques such as nearest neighbor and bilinear interpolation using qualitative as well as quantitative measures. The comparison shows the effectiveness of the proposed approach. The proposed approach can be used in applications where the memory, the transmission bandwidth and the camera cost are the main constraints.