Demosaicking based on Channel-Correlation Adaptive Dictionary Learning

Abstract

Image demosaicking is the problem of reconstructing color images from raw images captured by a digital camera which is covered by a color filter array (CFA). Sparse representation based demosaicking method achieves excellent performance on the commonly used Kodak PhotoCD dataset. However, it performs inferior on the IMAX dataset. In this paper, we analyze that the factor of the sparse representation based demosaicking methods perform different is channel-correlation. Hence, we propose a channel-correlation adaptive dictionary learning based demosaicking (CADLD). Different from the sparse representation based demosaicking methods that use one fixed dictionary, our method train a general dictionary on training datasets with varies channel-correlations. Then we learn a function matrix between the general dictionary and channel-correlations. For a raw data with a specific channel-correlation, we demosaick it adaptively to its channel-correlation through the function matrix. Experiments confirm the proposed method outperforms other sparse representation based demosaicking methods.