Digital cameras are widely used as digital image capture devices nowadays. Ideally three CCD sensors should be used to capture the three primary colors (red, green, blue), but CCDs are expensive. Thus single CCD with Bayer color filter array which samples one of three primary colors is commonly adopted in digital cameras. In order to reconstruct full color images, color demosaicking process is required to interpolate the two missing color samples from images acquired from the single CCD with Bayer filter. The color demosaicking is critical to the image quality....[ Read more ]

Digital cameras are widely used as digital image capture devices nowadays. Ideally three CCD sensors should be used to capture the three primary colors (red, green, blue), but CCDs are expensive. Thus single CCD with Bayer color filter array which samples one of three primary colors is commonly adopted in digital cameras. In order to reconstruct full color images, color demosaicking process is required to interpolate the two missing color samples from images acquired from the single CCD with Bayer filter. The color demosaicking is critical to the image quality.

In this thesis, three of our proposed directional color demosaicking algorithms which aim at choosing to interpolate along the edge instead of across the edge are presented. The first one is based on directional similarity in color difference spaces to adaptively choose interpolation direction (DSMCDS). Directional similarity is estimated by similarity measurement which provides hints on whether the direction for interpolation is along or across the edge. The second method is soft decision color demosaicking using our proposed simple pass filter which reconstructs better signals and vector selection process to consolidate the process of selecting appropriate interpolation direction (DDV). The third proposed approach is to categorize a direction for each pixel based on several considerations such as gradient or edge direction of neighbor pixels that exploits intra- and inter-channel correlation (DCat). Experimental results demonstrate that our three proposed algorithms significantly improve visual quality of reconstructed images and obtain much higher PSNR over bilinear interpolation.