Refining Color Constancy Under Single Illumination

For computer vision and image processing,color is an important cue.The color of the image is determined by the inherent characteristics of the object surface and the color of the light source.The change of the light source will make the color information unreliable.To solve this problem,the color constancy algorithm is dedicated to eliminate the influence of the non-standard light source or the light sensor in the imaging equipment on the image color,so that the color information in the image tends to be constant,which provides other computer vision related technologies stable information.However,this task is a pathological problem,which will be extremely unstable due to the change of unknown light source,object material and external imaging factors,which also makes the performance of the task algorithm encounter a bottleneck at this stage;in addition,datesets of single illumination color constancy all have the problem of label ambiguity,which is unavoidable.So at this stage,algorithms of this task are often low robustness and with poor generalization ability.To solve these problems,this paper mainly carried out the following research work:1.In this paper,a refinement strategy based on residual learning is proposed to modify the original light source estimation.The algorithm of color constancy based on the convolution neural network of statistics and learning is selected as the basic model to obtain the initial light source estimation value,and then the refinement optimization module is designed to fit the residual value between the initial light source estimation value and the Ground Truth,which alleviates the performance bottleneck of this kind of algorithm and solves the problem of low robustness due to the label ambiguity of datasets.In this paper,the feasibility of refinement optimization based on statistics and learning is verified.Experiments show that the learning based refinement module is more flexible,and the effect is better when using the learning based algorithm as the basic model.2.According to the verification results in 1,we take the structure of learning refinement improvement of convolutional neural network is as an important research object.To maximize the refinement effect of the structure,and make the information flow before and after the structure and assist learning,this paper discusses and designs the cascaded network structure and the cumulative multiplication loss function,and then proposes a new algorithm--Cascading Convolution Color Constancy(C~4).In addition,under the same experimental settings,the mean value and worst-25%of the angle error in the open dataset exceed the leading edge algorithm,breaking through the performance bottleneck of the task,and showing good robustness;at the same time,in the cross dataset experiment,the various measures of the angle error exceed the leading edge learning based algorithm,and the leading edge based on the unified algorithm Compared with the calculation method,it also has good comparability,which fully proves the good generalization of the algorithm.