Abstract—Biometric face Recognition is a new generation technology for identification and verification. Many techniques have been developed for this purpose in past two decades. A good face recognition technique must require unique, effective and efficient features from face images. Content based image retrieval (CBIR) can extract features that can be useful for face recognition. Recently, Gabor and curvelet texture features have been successfully used in image retrieval research. In this paper, we propose a novel face recognition method that uses texture features obtained by calculating mean and standard deviation of Gabor and curvelet transformed face images. PCA is then applied to the feature vectors instead of entire transformed images which traditional methods do. Using this process, we build four classifiers using mean and standard deviation calculated from Gabor and curvelet transformed face images. For identification purpose, a new matching strategy is proposed that checks goodness of four matching results of the classifiers. As we consider only mean or standard deviation features, the image representation has comparatively lower dimensions. Furthermore, our proposed method does not necessarily require all the input images to be of same resolution. We evaluate the proposed method using ORL and Yale face databases. The recognition results of the experiments show that our approach is significantly better than the conventional methods. Index Terms—Content based image retrieval (CBIR); gabor and curvelet transform; principal component analysis (PCA). I. INTRODUCTION We can recognize faces within a second that is a great demonstration of incredible human intelligence but this must be a very difficult problem for a computer [1]. Over the last two decades, researchers are making their best effort for achieving high performance rate in automated face recognition system. It is an interesting and challenging research field in biometrics and computer vision and is extensively used in security system, human-computer interface, surveillance and so on. To-date, researchers have proposed a number of face recognition approaches [1]. Most of them use Principal Component Analysis (PCA). PCA is a powerful tool for extracting some characteristic information of the dataset to identify pattern. It performs dimensionality reduction to input dataset and represents the whole face-space in lower dimension that contains most variance to highlight similarities and differences within the face-classes. PCA Manuscript received May 17, 2012; revised June 26, 2012. The authors are with the Department of Computer Science and Engineering Bangladesh University of Engineering and Technology(BUET) Dhaka- 1000, Bangladesh based approaches [3], [4] consider only raw pixel intensity of the face images. Furthermore, these approaches require that the input face images must be ideally aligned and under well-controlled illumination. To overcome these shortcomings, a number of approaches transform images by different filters like Gabor [5], [6] and curvelet [8], [9] and then apply PCA. However, they still use the raw pixel intensity of transformed images for face recognition. Such process inevitably implies high processing time and memory requirement for large dimensional feature vector. Due to face image is an unstructured array of pixels, we believe that the first step in understanding the face is to extract efficient and effective visual features from these pixels. There are thousands of publications in the literature describing color, texture, shape and spatial feature extraction for image retrieval [2]. But most of the approaches of face recognition consider only pixel intensity values of different transformation of images like Gabor [5], [6] or curvelet [8], [9] transformation. Raw pixel intensity values cannot be a good feature for face representation due to the curse of high dimensionality and semantic gap which is the difference between the high level interpretation of a face image and its low level pixel representation. The use of content based image retrieval (CBIR) could solve this problem [10], [11]. Content features like Gabor and curvelet texture can capture appropriate features having strong discriminative capability in image retrieval and semantic learning techniques. Being inspired from the success of image retrieval, we firstly use a new face image representation method that obtains texture features found by calculating the mean and standard deviation of the Gabor filtered or curvelet transformed images for face recognition. Although both Gabor and curvelet can capture face information efficiently, those features do not find same information because of having different spectrum in frequency domain. Therefore, it may happen that Gabor misses to capture some significant discriminative information where curvelet succeeds or vice versa. For this reason, we propose a new technique to combine those features efficiently that enable to function features independently. Final decision is generated by checking the goodness of performance of individual features. Moreover, our proposed method solves high dimensionality problem of image representation efficiently. Furthermore, a new flexibility is inserted in our approach that it can work properly although all input images have different dimensionality. In our proposed method, we first divide the face image into sub-images. After that, Gabor and curvelet texture features are extracted from the sub-images to feed into PCA and create four different classifiers. Each classifier can find an intermediate match using nearest neighbor Combination of Gabor and Curvelet Texture Features for Face Recognition Using Principal Component Analysis Shafin Rahman, Sheikh Motahar Naim, Abdullah Al Farooq, and Md. Monirul Islam 264 International Journal of Computer and Electrical Engineering, Vol. 4, No. 3, June 2012