Robust Face Recognition Using SPIDER Descriptor

Abstract

Real world face recognition systems require balancing of three concerns : compu- tational cost, robustness in changing environment and discriminative power. De- signing a face feature having small size would reduce computation cost but would also have minimal discriminative power and may not be stable in uncontrolled en- vironment and including more facial information to mitigate this challenge would increase the feature size and thus the computation cost. So, achieving a balance is a key aspect for every successful system. In our thesis we introduce an e ective appearance-based facial feature descrip- tor constructed with the new local texture pattern namely the Similarity Pattern of Image Directional Edge Response (SPIDER) for face representation and recog- nition. The proposed method encodes texture information of a center cell pixel by accumulating the directional edge response of all the pixels in the cell and then computing the dissimilarity measure of the local histogram of the center cell with its 8 neighbor cells using the Chi-square method. The dissimilarity values are then thresholded against a local average dissimilarity to generate a 8 bit-binary code which is assigned to the center pixel of the center cell. The distribution of the resulting SPIDER codes are then used as a face representation. To make the method compatible with real world systems a further step of dimension reduction is done to achieve fast classi cation time at the cost of a slight decrease in the accuracy. The e ectiveness of the proposed method has been evaluated using the FERET face image database using template matching and experimental results shows better performance of the SPIDER feature descriptor in comparison to other well-known appearance based feature extraction methods. i