Computational Color Constancy Based On Local Confidence And Contextual Information

The goal of computational color constancy is to eliminate the effects of illuminant variations in image scenes,which mainly involves two stages: illuminant estimation of the image scene,and image correction using the predicted illumination color.Some advanced methods sample local patches for illuminant estimation,and finally obtain global estimation results by simple pooling methods(e.g.average pooling or median pooling),where inaccurate local estimation is likely to affect the results of global estimation.Current patch sampling methods are mainly based on random sampling or grid sampling,which ignores the influence of the color value of each pixel on the illuminant estimation.In addition,behavioral studies suggest that contextual information in the image space is beneficial to the consistency of human color constancy.Existing methods often ignore the effect of contextual information.Based on the aforementioned issues,in this thesis,we conduct a study on computional color constancy based on local confidence and contextual information.The main contents are listed as follows.(1)We introduce the related work of computational color constancy,analyze the research significance of this task,and illustrate the fundamental of computatioanl color constancy.(2)To obtain the patches related to the color information of the light source,we design a patch sampling method based on bright and dark pixels,which can obtain the patches containing color gradient information.Furthermore,based on the VGG-16 network,we propose a patch-wise illuminant estimation network as the baseline model.Experiments verify the effectiveness of the proposed approach.(3)To effectively integrate local features of the sampled patches,considering the importance of different patches to illuminant estimation,we propose a confidence pooling method for global illuminant estimation.Based on the baseline model,we propose a shallow confidence network,where the confidence weight of the local patch is estimated and used to generate the estimation result of the global image.The experimental results show that the local confidence-based pooling significantly improves the accuracy of illuminant estimation compared with simple mean or median pooling methods.(4)Based on the influence of spatial contextual information on color constancy,we propose an illuminant estimation approach based on local context information to obtain accurate values of illuminant estimation,which consists of two-stream network(un-shared weights)with central patches and surrounding patches as inputs.In order to further improve the local illuminant estimation accuracy,based on the above network,we design a shallow refinement network by feeding the original patch stacked with the version corrected by the initial illuminant estimation result.The refined local illuminant estimation result is obtained by two successive stages.Experimental results verify the necessity of contextual information and refinement,and show that our approach achieves competitive performance on both benchmark datasets.