Face recognition under large illumination variations using homomorphic filtering in spatial domain Hossein Shahamat ⇑ , Ali A. Pouyan 1 Department of Computer Engineering and Information Technology, Shahrood University of Technology, No. 316, Daneshgah Ave., Shahrood 3619995161, Semnan, Iran article info Article history: Received 19 June 2013 Accepted 17 February 2014 Available online 4 March 2014 Keywords: Frequency domain filtering Butterworth high-pass filter Illumination normalization Face image preprocessing Face representation Homomorphic parameters Kernel function Reflectance component abstract This paper proposes a homomorphic filtering in spatial domain for reducing of illumination effects in face recognition systems. Also, in this research a simple kernel of homomorphic filter is proposed. Application of this method causes considerable reduction in computational time in the preprocessing step. When a new face image with an arbitrary illumination is given, the homomorphic filter is applied and its reflec- tance component is extracted. Then the reflectance component is divided into several local regions and histograms of each local region are extracted using multi-resolution uniform local Gabor binary patterns (MULGBP). These histograms are combined for obtaining the overall histogram of the images. Finally, for face recognition, a simple histogram matching process is performed between new face image histogram and the gallery images histogram. The results show that the proposed method is robust for large illumi- nation variation with a reasonable computational complexity. Ó 2014 Elsevier Inc. All rights reserved. 1. Introduction Face recognition is increasingly investigated for personal secu- rity and access control applications. A challenge in face recognition is finding efficient and discriminative facial appearance descriptors that can counteract large variations in illumination, pose, facial expression, aging, partial occlusions and other changes [1]. In con- trast with other biometric analysis, face is generally regarded as the most accepted component for users, since it provides a friendly and convenient method of identification. Many researchers have investigated face recognition and several algorithms have been proposed in the last two decades. While majority of them work well under controlled environments they exhibit poor performance when face images are captured under uncontrolled conditions. Several face image preprocessing methods have been proposed to cope with illumination changes. Histogram equalization (HEQ) [2] is one of the most useful contrast enhancement schemes. How- ever, since HEQ technique only enhances the contrast of global im- age in spatial domain, it does not particularly consider the details involved in face images. Wang et al. [3] proposed the self-quotient image (SQI), which is defined as the ratio of the input image and its smooth versions. It is based on the Quotient Image method [4] to achieve lighting invariant. The wavelet-based illumination normal- ization method [5] applies histogram equalization to the low fre- quency and accentuate the high frequency coefficients. Jobson et al. [6] proposed the multi scale retinex (MSR) method. It cancels much of low frequency information through dividing the image by a smoothed version of itself. Xie et al. [7] decomposed a face image into large- and small-scale features, normalized illumination sepa- rately on both features, and finally recovered a normal illumination image through combining both features together. Tan and Triggs [8] proposed an integrative framework that combines the strengths of robust illumination normalization, local texture based face representations, distance transform based matching, kernel- based feature extraction and multiple feature fusion. Using We- ber’s Law, Wang in [9] considers the ratio between local difference and the center degree as a kind of illumination invariant compo- nent. These methods are quite effective but their ability to handle extreme uneven illumination variations remains limited. A novel preprocessing method based on statistical bilinear model [10] is proposed by Jun et al. [11]. This method transforms input face im- age into sixteen face images exhibiting different illuminations. It yields high recognition rates but its computational complexity is not reasonable. In this paper, illumination effects are reduced using homomor- phic filter in spatial domain, instead transforming an image into sixteen images using bilinear model. It reduces total face recogni- tion time. Furthermore, the proposed method is applied on two http://dx.doi.org/10.1016/j.jvcir.2014.02.007 1047-3203/Ó 2014 Elsevier Inc. All rights reserved. ⇑ Corresponding author. Fax: +98 273 3391600. E-mail addresses: Shahamat@shahroodut.ac.ir, Hossein_Shahamat@yahoo.com (H. Shahamat), apouyan@shahroodut.ac.ir (A.A. Pouyan). 1 Fax: +98 273 3391600. J. Vis. Commun. Image R. 25 (2014) 970–977 Contents lists available at ScienceDirect J. Vis. Commun. Image R. journal homepage: www.elsevier.com/locate/jvci