Demosaicing and Zooming for Color Filter Array via Residual Image Reconstruction Demosaicing and Zooming for Color Filter Array via Residual Image Reconstruction

A method of demosaicing and zooming for color filter array (CFA) via residual image reconstruction is presented. Given the demosaicing and zooming methods, a residual image between the ‘genuine’ image and an initial demosaiced and zoomed image is reconstructed using a dictionary and sparse coding. The ‘genuine’ image has richer edges and details than that of the initially obtained image. A generic dictionary is generated by learning a large set of training images and their sparse representations. An adaptive dictionary specific to the input CFA is also learned. Using the adaptive dictionary, an initial residual image is reconstructed, and an optimal scaled version of the residual image is then calculated depending on the input CFA. The optimal scaled residual image is added back to the initially obtained image to get the final result. Experimental results show the performance in terms of PSNR and visual quality. Keywords-Demosaicing; Zooming; Residual image reconstruction; Sparse representation INTRODUCTION In most resource-constrained digital image/video capture devices, only single chip and color filter array (CFA) are used [1]. Often, full color and enlarged image produced from a CFA are both needed. Demosaicing or spectral interpolation is performed to get a full color image and zooming or spatial interpolation is performed to get an enlarged image. There are three categories of schemes to obtain a full color enlarged image in general: the CFA is demosaiced first and then zoomed [2]; the CFA is zoomed first with the same pattern as the input CFA and then demosaiced[3,4]; demosaicing and zooming are implemented simultaneously. While plenty of both demosaicing and zooming techniques have been proposed in literature and their combinations have generated satisfied results [5, 6], there are still lots of improvement spaces. For instance, as stated in [2], after being demosaiced, whether individual channel or three channels altogether are zoomed, the color artifacts generated during demosaicing will be amplified during zooming. In our proposed framework, first, a full color and enlarged image is obtained by using certain demosaicing and zooming method, and then we assume that there exist a ‘genuine’ image having richer edges and details. Our goal is to search edges and details lost by the current enlarged full color image and the lost edges and details compose a residual image. We make use of sparse representation and sparse coding to reconstruct the residual image. After an initial residual image is reconstructed, an optimal scale factor is calculated to multiply with the initial residual image to get appropriate values. It is declared that in principal, arbitrary demosaicing and zooming techniques or their combinations are allowed to get the initial full color and enlarged image; however, to achieve more satisfied results, better demosaicing and zooming techniques are expected. The rest of the paper is structured as follows: section 2 outlines and discusses our proposed methods in detail; section 3 provides experimental results; section 4 concludes the paper.