Haar Local Binary Pattern Feature for Fast Illumination Invariant Face Detection

Face detection systems based on boosted Haar features in a cascade architecture [3] perform very well both in terms of accuracy and speed. However, they are vulnerable to variations in illumination conditions. On the other hand, face detection systems based on the Local Binary Pattern (LBP) [2] or its variants [1] show robustness to illumination conditions but do not take advantage of the richness of the Haar feature set in efficiently modelling faces. In this paper, we propose a face detection system based on a new type of feature called the Haar Local Binary Pattern (HLBP) feature which combines the advantages of both Haar and LBP and is boosted using AdaBoost as in [3]. This feature compares the LBP label counts in two adjacent image subregions similar to Haar masks, i.e. it indicates whether the number of times a particular LBP label occurs in one region is greater or lesser than the number of times it occurs in another region, offset by a certain threshold. They capture the region-specific variations of local texture patterns which is hypothesized to remain relatively stable across variations in illumination conditions. Experiments show that our features more robust to illumination variations, compared to Haar and LBP individually. Let the input be an N×M gray-level image, represented as an N×M matrix I. Firstly, the LBP image ILBP [2] is calculated from the original input image I using LBP4,1, (ref. Fig. 1) found to be the optimal operator in our case. Next, the Integral Histogram set {IH k } 16 k=1 [4] of the LBP image ILBP is calculated. The individual pixels IH k (x,y) of the k-th Integral Histogram IH k is calculated as the number of pixels above and to the left of the pixel (x,y) in the LBP image ILBP which have a label k, i.e. IH k (x,y) = ∑u≤x,v≤y δk(u,v), where δk(u,v) = 1 if the label of the pixel at location (u,v) in the LBP image ILBP is k, and is zero otherwise. Using the following pair of references, for all k ∈ {1,16} :