7.3-5 Abstract—This paper presents a novel local feature descriptor, the Local Directional Pattern (LDP), for recognizing human face. A LDP feature is obtained by computing the edge response values in all eight directions at each pixel position and generating a code from the relative strength magnitude. Each face is represented as a collection of LDP codes for the recognition process. I. INTRODUCTION The recent advancement of software and hardware technology has created more demand for personalized interaction with consumer products. Popular method for achieving this by identifying the users through human face recognition and enabling appropriate services, such as personalized TV program. A recent survey on face recognition [1] explicates the significance and progress of this research domain [2]. However, robust face recognition system in uncontrolled environment is still a major challenge. Finding efficient facial features to represent the face appearance is the most critical aspect in face recognition. Facial features fall into two classes – global feature and local feature. In global feature extraction process, the whole image is taken into account, but local feature considers only the local region within the image. Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), 2D PCA etc., generates global features that have been widely used. Although, the global features are popular in face recognition, their performances deteriorate in changing environment. Hence local features are gaining more attention for their robustness in uncontrolled environment. Local Features Analysis (LFA), Elastic Bunch Graph Match, Dynamic Link Architecture (DLA) are popular among the local methods for locating the local face features. Ahonen et al. [3] proposed Local Binary Pattern (LBP) that provides an illumination invariant description of face image. However, the existing methods still suffer much from non-monotonic illumination variation, random noise, and change in pose, age, expression. This paper describes a Local Directional Pattern (LDP) which overcomes the drawbacks of LBP and is more robust in recognizing face. The proposed new local descriptor LDP considers the edge response values in all different directions instead of surrounding neighboring pixel intensities like LBP. This provides more consistency in the presence of noise; since edge response magnitude is more stable than pixel intensity. II. LOCAL BINARY PATTERN (LBP) The LBP operator, a gray-scale invariant texture primitive, has gained significant popularity for describing texture of an image [4]. It labels each pixel of an image by thresholding its * Corresponding author. P-neighbor values with the center value and converts the result into a binary number by using (1). 1 , 0 1 0 ( , ) ( )2 , () 0 0 P p PR c c p c p x LBP x y sg g sx x - = ≥  = - =  <   (1) where g c denotes the gray value of the center pixel (x c , y c ) and g p corresponds to the gray values of P equally spaced pixels on the circumference of a circle with radius R. Ojala et al. [4] also observed that in significant image area certain local binary patterns appear frequently. These patterns are named as “uniform LBP” as they contain very few transitions from 0 to 1 or 1 to 0 in circular bit sequence. Ahonen et al. [3] used this variant of LBP patterns which have at most two transitions (LBP u2 ) for their face recognition. This variant of LBP is still sensitive to random noise and non- monotonic illumination variation. III. LOCAL DIRECTIONAL PATTERN (LDP) The proposed Local Directional Pattern (LDP) is an eight bit binary code assigned to each pixel of an input image. This pattern is calculated by comparing the relative edge response value of a pixel in different directions. For this purpose, we calculate eight directional edge response value of a particular pixel using Kirsch masks in eight different orientations (M 0 ~M 7 ) centered on its own position. These masks are shown in the fig. 1. 3 35 3 0 5 3 35 - -     -     - -   3 5 5 3 0 5 3 3 3 -     -     - - -   5 5 5 3 0 3 3 3 3     - -     - - -   5 5 3 5 0 3 3 3 3 -     -     - - -   East (M0) North East (M1) North (M2) North West (M3) 5 3 3 5 0 3 5 3 3 - -     -     - -   3 3 3 5 0 3 5 5 3 - - -     -     -   3 3 3 3 0 3 5 5 5 - - -     - -       3 3 3 3 0 5 3 5 5 - - -     -     -   West (M4) South West (M5) South (M6) South East (M7) Fig. 1: Kirsch edge response masks in eight directions Applying eight masks, we obtain eight edge response value m 0 , m 1 , …, m 7 , each representing the edge significance in its respective direction. The response values are not equally important in all directions. The presence of corner or edge show high response values in particular directions. We are interested to know the k most prominent directions in order to generate the LDP. Hence, we find the top k values | m j | and set them to 1. The other (8-k) bit of 8- bit LDP pattern is set to 0. ( ) 0 1 7 , : ( 1) 0 7 ( ); { , ,..., } i i th C f xy c if i and m where k M M m m m ψ ψ = = ≤ ≤ ≥     = = (2) The LDP code produces more stable pattern in presence of noise. For instance, fig. 2 shows an original image and the Local Directional Pattern (LDP) for Face Recognition Taskeed Jabid, Md. Hasanul Kabir, and Oksam Chae*, Member, IEEE Kyung Hee University, South Korea. 978-1-4244-4316-1/10/$25.00 ©2010 IEEE