K-means and particle swarm optimization based color constancy of images

Abstract

Color constancy is a fundamental requirement for many image processing and computer vision applications. Since color constancy is an under constrained problem, the existing methods are based on some assumptions and hence no method exists that works for all types of images. The proposed approach uses Particle Swarm Optimization (PSO) algorithm to combine different existing algorithms in an optimal way so that a single method works for almost all types of images. The combination of existing algorithms is done in weighted proportion, where each existing algorithm has an associated weight, which varies with the type of image. These weights are learned during the training phase. In the training phase, similar types of image are clustered using K-Mean clustering algorithm. The clustering is performed over Weibull parameters and for each cluster the optimal weights are obtained using PSO. Once the system is trained, given any input image the system can correct the image by grouping the image into one of the clusters based on similarity measure, and then applying the optimal weights corresponding to that group/cluster (obtained during training). The median angular error criterion is used to compare the results of the proposed approach with some of the existing color constancy methods. Obtained results show the effectiveness of the proposed approach compared with other considered approaches.