A New Fuzzy-Based Color Reflectance Space with Selective Feature Extraction: Face Recognition Under Diversified Illuminations.

Effective face recognition becomes challenging when significant facial features diminish due to variations in light energy distribution incident on images. Many existing face recognition techniques have neglected color information while annulling low-frequency-based illumination effects due to which their performance is compromised during learning. But, low-frequency color textural features are desirable in accelerating learning tasks along with high-frequency edge features. To cope with light variations on color images, this work aims to propose a new light-variation normalization framework and feature extraction for face recognition. The main contributions are: a new framework for annulling light variations on images using reflectance components of color pixels in RGB plane proposed as color reflectance space. This combines with local color intensification and enhanced contrast based on redistribution of dynamic pixel-intensity range in proposed reflectance space using local fuzzy contrast adaptation. Feature extraction is then performed in double-density dual-tree complex wavelet transform domain on normalized images. These robust features are efficient in facilitating classification process. The designed face recognition system is competent in handling light variations; however, pose invariance is not addressed in this work. Efficacy of proposed face recognition is established on color face databases, namely CMU-PIE and AR, where 99.66% and 100% recognition rates are achieved, respectively. The performance is also validated on grayscale images with extreme light variations (Subset-5) of Extended Yale-B database where 99.06% is achieved. It is concluded that comparative assessment of attained results by proposed face recognition under diverse illuminations is superior than other existing methods. [ABSTRACT FROM AUTHOR]