Research On Algorithms For Face Recognition Under Complex Illumination Conditions

Using biometric technology to identify individual identity has become the primary way of secure authentication. The face recognition is one of the topics in the most commonly used means of identification, but also one of the most popular pattern recognition research. The face recognition technology, after nearly 50 years of development, obtains a wealth of experience and algorithms. However, under uncontrolled conditions, especially illumination, face recognition is still a challenging research task. The face recognition algorithms under complex illumination conditions are studied. The focus of my work is the image preprocessing and extracting face features under complex illumination variations.With in-depth analysis of the face image preprocessing algorithms, the study focuses on the de-noising processing, geometric normalization and light preprocessing methods. Because a face images in the transmission and storage process is disturbed by pulsed interference seriously, and median filtering proved by experiments in addressing the face noise has a great advantage, so this article will use the median filter to solve the noise issue. If the standard of the original face images is different, it would seriously affect the accuracy of face recognition. A uniform standard for the face recognition system is very important. This paper summarizes previous knowledge and gives the normalization standards of face images. Several common illumination pretreatment methods are studied. In practical engineering applications, the speed and effectiveness are often the two factors that researchers must consider. For PS with Different-of-Gaussian-based in the speed has a great advantage, but the effectiveness meets the requirement, so this method is selected as the illumination pretreatment approach in my article.The application of local binary pattern and its variants in face recognition are studied. Since local texture features can well describe the structural information of a face, while the robustness of the pose and illumination is high, so in this paper three representative algorithms are studied, which are LBP, LTP and MB-LBP. Since MB-LBP feature dimension is large and the dimensionality reduction algorithm of LBP and LTP is no longer applicable to MB-LBP, this article will introduce another dimensionality reduction method, which is linear discriminant analysis. Linear discriminant analysis is commonly used in face recognition. This method can reduce the number of dimensions and obtain more discriminative features in the same time. However, this method usually has a small sample problem, thus the principal component analysis for the original characteristics is needed before the linear discriminant analysis. In order to maximize the extraction of face image feature, so the multiresolution technique is studied. For wavelet transform has the characteristics of the multi-resolution and conform to the mechanism of thuman eyes from coarse to fine recognition process, wavelet transform is brought into MB-LBP.Feature extraction is the key step in face recognition. The extracted features must have strong robustness to illumination, expression, posture change. MB-LBP with WT gets a good recognition rate with 98.31% in Extented YaleB face database.