Comparison of demosaicking Methods for Color Information Extraction

Most digital color cameras are based on a single CCD or CMOS sensor combined with a color filter array (CFA): each pixel measures only one of the RGB colors. The most popular CFA is the Bayer CFA (Bayer, 1976) shown in fig. 1. Demosaicking algorithms interpolate the sparsely sampled color information to obtain a full resolution image, i.e. three color values per pixel. Many demosaicking algorithms were designed. However, even recent methods are prone to interpolation errors or artifacts, especially near edges. The most common artifacts are shown in fig. 2: “zipper” effects, wrong colors and wrong saturation of colored details. These artifacts are influenced by the sampled channel (R, G or B) and by the image content, like e.g. the orientation of the nearby edges. As a consequence, the same scene point may get a very different color value after a camera movement. This raises the two questions whether images acquired with a single chip camera can be used to extract reliable color information for further computer or robot vision tasks, and which demosaicking method suits best. Hence, this paper provides an overview and a detailed comparison of several state of the art and several recent demosaicking algorithms to enable the reader to choose the most appropriate demosaicking algorithm for his application. The previous comparisons between demosaicking methods, like the ones in (Lu & Tan, 2003; Ramanath et al., 2002), aimed at visually pleasing images. As a consequence, their evaluation criteria were, in addition to Mean Square Error (MSE) in RGB space, visual inspection and measures based on human perception like * ab E ∆ (Lu & Tan, 2003; Ramanath et al., 2002). In computer vision tasks, separation between intensity and chrominance is widely used to increase robustness to illumination changes (Funt et al., 1998; Gevers & Smeulders, 1999). However, none of the previously used criteria can evaluate chrominance quality. For that reason, a detailed analysis using MSE in typical color spaces (HSI, Irb and YUV) is provided here. In addition, performance differences in colored, textured and homogeneous areas are emphasized. After an overview of the existing demosaicking algorithms, the methods chosen for comparison are introduced in section 2. The comparison framework and the results are presented in section 3. A conclusion is given in section 4. G R G R G