Improvement of Bayer-pattern demosaicking with dictionary learning algorithm

Demosaicking is important for the quality of digital images in resource-constrained single chip devices. This paper presents an improved dictionary learning-based color demosaicking algorithm. Firstly, an initial interpolation is applied to the R , B channel by Local Directional Interpolation (LDI) and fused by analysis the joint distribution of the gradient. Gaussian Mixture Model (GMM)-based clustering is used to classify dictionary image into different classes. The Principal Component Analysis (PCA) is performed on these classes to choose the principal components for the dictionary construction. And then, dictionary learning is applied to obtain the interpolated Ĝ and the lost R̂ and B̂ are interpolated by the help of the reconstructed Ĝ, accordingly. Since R̂, Ĝ andB̂ of the given pixels are better represented, the whole image can be reconstructed accurately. Taking McMaster color image dataset as dictionary, standard image and image from DALSA CMOS camera are used for effect evaluation of the demosaicking algorithm. Experimental results prove that the proposed algorithm outperforms some state-of-the-art demosaicking methods both in PSNR measure and visual quality.