Improvements in illumination compensation for face recognition under noncontrolled illumination conditions

Abstract

Face recognition depends strongly on illumination conditions, especially in non-controlled scenarios where face illumination is not homogeneous. For this reason, illumination compensation is crucial in this task. Several methods for illumination compensation have been developed and tested on the face recognition task using international available face databases. Among the methods with best results are the Discrete Cosine Transform (DCT), Local Normalization (LN) and Self-Quotient Image (SQI). Most of these methods have been applied with great success in face recognition using a principal component classifier (PCA). In the last decade, Local Matching Gabor (LMG) classifiers have shown great success in face classification relative to other classifiers. In all cases, the illumination compensation methods improve the face recognition rates in unevenly illuminated images, but affect negatively in some well illuminated images. The aim of this thesis is to propose improvements to the current illumination compensation methods to obtain improved face recognition rates under different illumination conditions. Using genetic algorithms (GAs), parameters of the SQI method were selected to improve face recognition. The parameters optimized by the GA were: the fraction of the mean value within the region for the SQI, selection of Arctangent, Sigmoid, Hyperbolic Tangent or Minimum functions to eliminate noise, and the weight values of each filter are selected within a range between 0 and 1. The results obtained after using the proposed method were compared to those with no illumination compensation and to those previously published for SQI method. Four internationally available face databases were used: Yale B, CMU PIE, AR, Color FERET (grayscaled), where the first three contain face images with significant changes in illumination conditions, and the fourth one contains face images with small changes in illumination conditions. The proposed method performed better than SQI in images with non-homogeneous illumination. In the same way, GAs were used to optimize parameters of the modified LN and SQI methods in cascade for illumination compensation to improve face recognition. The main novelty of this proposed method is that it applies to non-homogeneous as well as homogeneous illumination conditions. The results were compared to those of the best illumination compensation methods published in the literature, obtaining 100% recognition on faces with non-homogeneous illumination and significantly better results than other methods with homogeneous illumination. Also, the DCT, LN, and SQI illumination compensation methods were optimized using GAs to be used with the LMG face classifier. Results were tested on the FERET international face database. Results show that face recognition can be significantly improved by modified versions of the current illumination compensation methods. The best results are obtained with the optimized LN method which yields a 31% reduction in the total number of errors in the FERET database. Finally, an extension of the LN method using Kolmogorov-Nagumo-based statistics was proposed to improve face recognition. The proposed method is a more general framework for illumination normalization and it was showed that LN is a particular case of this framework. The proposed method was assessed using two different classifiers, PCA and LMG, on the standard face databases Extended Yale B, AR and Gray FERET. The proposed method reached significantly better results than those previously published for other versions of LN on the same databases.