Uneven Illumination Compensation for Unconstrained Face Recognition Using LBP

Abstract

Face recognition results are degraded when the test image is having nonuniform illumination variations. Due to the lighting variations, different expressions and occlusions the facial appearance of the face changes dramatically such that recognition by the methods is quite a difficult task. Out of which the problems due to illumination variations such as shadows, under lighting, over lighting in the face are the crucial problems which are to be overcome to achieve the satisfactory results for an automatic face recognition system. In this paper, we propose an efficient illumination compensation method based on frequency analysis and multi-resolution analysis. The effect of uneven lighting variations of the test image is efficiently and effectively removed by applying the LBP image to modify the magnitude information in the frequency domain. The magnitudes of the LBP image compensate the distorted magnitudes of the original image, caused by the light variations, as well as add some structural information to the restored image. To examine the efficacy of our developed method, histogram equalization based normalization and Fourier based normalization methods are compared with the proposed method. An extensive simulation study has been carried out to measure the effectiveness of our proposed method on the images containing extreme illumination variations. For this purpose, we have used the Extended Yale-B face database. From the results, it is found that our proposed method outperforms other methods and also it works better for extremely poor illuminated images.