Kernel Feature Extraction Approach For Color Image Recognition

Color images play an important role in current pattern recognition and machine learning. Color images have widely application for object detection, tracking and recognition. Compared to the grayscale images, color images possess much more useful information for improving image recognition performance, so color images attract more attention. The key of color image recognition technique is how to effectively utilize the complementary information between color components and remove their redundancy. Multi-set canonical correlation analysis is an important technique of color image recognition, which is used for color image, maximizes correlation coefficient between the color image components and extracts canonical correlation feature.Kernel method is an effective technique for extracting nonlinear feature. Recently kernel-based nonlinear analysis has been given more attention in the pattern recognition, because the kernel trick can efficiently construct nonlinear relations of the input data sets in an implicit feature space obtained by nonlinear kernel mapping. The kernel trick does not need to compute the inner products explicitly in the feature space, so it reduces the complexity of the algorithm in a certain degree.In this paper, we integrate color image recognition with kernel method and propose two novel color image recognition approaches:color image kernel canonical correlation analysis (CIKCCA) and color image kernel holistic orthogonal analysis (CIKHOA). Both approaches project the color image sets to nonlinear high dimensional kernel space using kernel trick. Color image kernel canonical correlation analysis is based on the theory of multi-set canonical correlation analysis and extracts canonical correlation feature between the color image components. Then fuse the features of the color image components in the feature level, which are used for classification and recognition. The basic idea of color image kernel holistic orthogonal analysis is that in the nonlinear high dimensional kernel space it can in turn extract discriminant feature of color image components and make the discriminant vector sets mutually orthogonal. CIKHOA reduces the correlation among the discriminant vector sets in the feature level and fuses the feature for classification and recognition. Experimental results on the AR, FRGC-v2, PolyU public color image databases demonstrate that the two proposed approaches acquire better recognition performance than other color recognition methods.