Color compensation via color-flow representation and eigenspace manifold learning for robust color-invariant face recognition

摘要

This paper presents a novel color compensation algorithm for improved color-based face recognition (FR) under non-controlled illuminants. Differing from the previous approaches, the underlying idea behind our method is to take advantage of a pair of the probe and gallery face images available to a typical color FR framework. To this end, a new and novel approach for deriving color-flow value representation and constructing color-flow eigenspace manifold learning has been developed to reliably estimate varying illuminations imposed on probe images. In addition, a sophisticated reconstruction solution has been developed to generate color compensated probe images whose illumination condition becomes much similar to the canonical illumination state of gallery images. Comprehensive and comparative experiments have been performed to demonstrate the effectiveness of our color compensation. For this, both quantitative and qualitative assessments of our method over other state-of-the-art color compensation techniques have been performed. Results show that our color compensation outperforms other color compensation techniques in terms of compensating color face images with non-linear colored-light and illumination cast shadow. Also, it can be shown that our novel framework that incorporates the proposed color compensation into recently developed color FR algorithms (as premilinary step) can significantly improve FR performances for challenging illuminant face images (with performance gains of up to 26 % for particular cases). The reported work provides a new insight into the merits of color compensation methods, as well as their role in dealing with severe illumination changes in color FR.