Demosaicing Research Based On The Structure Remain

Due to the cost and the complexity of physical structure, most of the imaging device uses a single chip CCD as the image sensor, which is covered with a layer of color filter array. The array only measures one color component in the RGB three primary colors. In order to estimate the other two missing color components, image must be interpolated to get the full color image. This process is known as demosaicing or CFA interpolation. As a foundation of image recognition, image segmentation, CFA interpolation has been widely concerned.In this paper, we first introduce the existing CFA interpolation algorithms, then, in order to preserve the structure information, we propose the CFA interpolation algorithms for the Bayer mode, The main work of this paper includes:1:After analyzing the shortcoming of the existing CFA interpolation algorithms, we propose a new CFA interpolation algorithm, which is based on the super-resolution reconstruction of the residual plane. In this way, color image demosaicing is converted into a image super resolution problem, firstly, a pair of high and low resolution dictionaries are learned, which describe the high and low frequency details of the G component of feature plane and residual plane,respectively. Then based on the sparse representation theory, the residual planes are added back to the low resolution images to obtain the final reconstruction image. With the aid of the rich details of high resolution images, the demosaiced image have more edge and texture information.2:For the R/B component interpolation, an improved directional weight interpolation method is proposed. Based on the directional interpolation, we calculate the unknown pixel of four direction gradient absolute value and the square, then the residual values are weighted estimated.3:By comparing existing refinement methods, we propose a regression judgment function based resembling demsoaicing algorithm. The edge similarity measure for the color difference plane is presented. Then combined with the gradient information, a linear regression function is obtained. In this way, by resembling the different two demosaicing methods, we achieve an effective integration refinement methods.