The Research Of Face Recognition Under Unconstrained Condition Via LBP And Deep Belief Network

Along with the research of half century, face recognition becomes one of the most active topics in pattern recognition, and makes great headway. Some face recognition systems can achieve quietly high performance, even can satisfy the requirement of practical use. But when the environment of the systems become complex and volatile, the performance decrease rapidly. For instance, face recognition under unconstrained condition considers illumination, pose, occlusion and so on. It is more close to real world than constrained condition and has more practical application value. Because of the variability and uncertainty, face recognition systems are lack of stability. Among the influencing factors, unconstrained illumination is one of the main bottlenecks in restricted face recognition development. At the same time the existing feature extraction algorithms can’t get strongly discriminative features, and the expression of features relies on the artificial selection. But in practice, researchers don’t known how to choose and express the features correctly.Aim to deal with the problem existing in the research of face recognition under unconstrained condition and illumination condition, in this paper we have discussed local binary pattern, wavelet transform and deep belief network which is a typical representation of deep learning nets. The main works of this paper are as follows:(1) To solve the problem of being lack of efficient feature representation in unconstrained illumination, face recognition based on wavelet and LBP in logarithm has been studied. This paper utilizes the superior high-pass filtering characteristics of wavelet in logarithm, thus accelerates illumination compensation of face images. Then LBP textural features are extracted from face images to further reduce the effects of illumination, and improve the ability of expression and discrimination of features. Experimental results on Yale-B and CMU-PIE face database show that our algorithm can be robust to unconstrained illumination and achieve better recognition rate.(2) To solve the problems of face images which can be easily affected by composite factors, and feature extraction relying on too much active factors involved, deep learning is proposed to automatically learn and extract features of face images for high-precision face recognition under unconstrained condition. Deep learning simulates the architectural depth of brain, and learns structural and multi-level features from bottom to up without relying on artificial feature selection. The extracted features benefits to visualization and characterizes the substantial character of data. Experimental results indicate that features extracted by deep learning are more suitable for face recognition under unconstrained condition. (3) To deal with the insufficient of DBN in face recognition under unconstrained condition, the novel operation of treating LBP textural features as the inputs of deep network is proposed. When pixel-level face images are affected by illumination, expression and so on. adverse feature descriptions may be learned by deep network. Meanwhile. DBN ignores the two-dimensional structural information of face images, which undoubtedly will be detrimental to excavate apparent information of face images. In this paper, LBP textural features are treated as the inputs of DBN for reducing the influences of external environment, and avoiding too much structural features lost. The experimental results on LFW (Labeled Face in the Wild).Yale and Yale-B face database show that more discriminative features can be obtained and high-precision face recognition under unconstrained condition can be achieved through the combination of LBP with DBN.