Research On The Color Image Demosaicing Algorithm Based On Residual Interpolation

Given the physical implementation and cost issues,most digital cameras use only one sensor for color images.The surface of the sensor is covered with a color filter array.The light passes through the color filter array to receive only one color component of the three primary colors of red,green and blue.To get a complete color image,we need to estimate the two missing color components,the process of image interpolation is called demosaicing.Demosaicing is the prerequisite of image recognition,object detection and other follow-up processing,so the study of effective demosaicing algorithm has important practical value.In this paper,based on the existing interpolation algorithm,a convolution neural network demosaicing algorithm based on residual interpolation is proposed.The innovation of this article is as follows:(1)When interpolating the green channel,we divide the image edge into more categories.Besides the traditional horizontal and vertical edges,we also include four kinds of edges.If there is no edge in both horizontal and vertical directions,it will determine whether there is an edge in the inclined direction.The edge detection operator in the inclined direction is composed of the gradient of the green channel and the two derivative of the red blue channel,which fully exploit the information of each channel.The interpolation direction is more accurate after the edge classification is refined.(2)The interpolation formula of the inclined direction is put forward.After a more detailed division of the edges,interpolation is done according to the detected edge direction.There is a very mature interpolation formula in the traditional horizontal and vertical directions.This paper focuses on the oblique direction interpolation formula.If the edge direction has been determined,according to the principle of channel correlation,a reasonable hypothesis can be put forward,that is,the color difference along the edge direction is stable.Based on this,we deduce the interpolation formula of oblique direction based on channel correlation,and interpolate accurately in the tilting direction.(3)After the green plane is interpolated temporarily,we can observe that the resolution of the preliminary result is not high enough.Based on this observation,our paper using convolutional neural network to modify the result of demosaicing,with the preliminary demosaicing result as the convolutional neural network input,and a full color image as the network label.The network was trained to study the mapping between the demosaicing result and original image,so as to realize the enhancement of the resolution.(4)A post processing step based on the color saturation of the mixed chromatic aberration plane is proposed.The experiments show that the higher the color saturation is,the weaker the correlation between the channels and the more unstable the chromatic aberration plane.Therefore,this paper proposes a method of mixing color difference plane based on color saturation and geometric distance,and the image recovered in high frequency region is better.