Face Recognition Based On Local Descriptor

With the rapid development of computer and network technology, the society becomes more information-rich. The traditional authentication model cannot satisfy the requirements of high-speed information and the biometric identification technology has aroused wide concern in various fields. As one of the most successful applications of image recognition field, face recognition has recently received significant attention, especially during the past few years. A good object representation or object descriptor is one of the key issues for a well-designed face recognition system.The Local Binary Pattern(LBP) is an efficient local feature descriptor and is becoming a popular technique for face representation as well as for image representation in general for its several advantages. This dissertation is mainly concerned with the original LBP and variants, Based on these algorithms, this paper first present an convex-concave pattern and propose a new method named as a Local Ternary Derivative Pattern(LTDP). The author’s major research work are outlined as follows:1. The original LBP and its variants have been studied. The local binary coded operator’s strengths and weaknesses have been investigated according to its definition. This paper establish a convex-concave model and propose a new local descriptor operators which named Local Ternary Derivative Pattern(LTDP). Based on the derivative of each pixel constructed by computing the values between the referenced pixel and its adjacent pixels with diverse distances from different directions, the convex-concave representation of the referenced pixels is generated to provide the 2D structure of micropatterns. Compared with other local algorithm, the LTDP take account of the local spatial information and global direction variations at the same time and extract more discriminative information. On the other hand, with an adaptive function, the LTDP perform more robust in the case for complex environments.2. Experiments on an extensive set of face databases, Extended Yale B, CMU PIE, FERET and CAS-PEAL databases are conducted to evaluate the comparative performances of LTDP, LBP and LDP(Local Derivative Pattern). The experimentalresults show that the LTDP performs much better than LBP and LDP and also confirm the validity of convex-concave pattern.3. In addition, face image have to face the problems of fuzzy background and uneven illumination due to the presence of differences in image acquisition environment. In order to illuminate the impact of the background, we proposed the image preprocessing scheme which named Do G filter and Gamma Correction, and combined the scheme with the proposed method. The comparative experiments on the set of face databases, Extended Yale B, CMU PIE, FERET and CAS-PEAL databases, show that the recognition performances of these local descriptors are improved and the LTDP consistently performs better than LBP and LDP with the preprocessing scheme. Another interesting fact is that the fluctuation of the identification accuracy of the LTDP with or without the preprocessing procedure is smaller than those of the two other methods. This is believed that the coding function of the LTDP inherently alleviates, to certain extend, the illumination variations and noise sensitivity problem in the gray-level images.