Robust face representation and recognition under low resolution and difficult lighting conditions

Abstract

Listen

Translate article

This dissertation focuses on different aspects of face image analysis for accurate face recognition under low resolution and poor lighting conditions. A novel resolution enhancement technique is proposed for enhancing a low resolution face image into a high resolution image for better visualization and improved feature extraction, especially in a video surveillance environment. This method performs kernel regression and component feature learning in local neighborhood of the face images. It uses directional Fourier phase feature component to adaptively lean the regression kernel based on local covariance to estimate the high resolution image. For each patch in the neighborhood, four directional variances are estimated to adapt the interpolated pixels. A Modified Local Binary Pattern (MLBP) methodology for feature extraction is proposed to obtain robust face recognition under varying lighting conditions. Original LBP operator compares pixels in a local neighborhood with the center pixel and converts the resultant binary string to 8-bit integer value. So, it is less effective under difficult lighting conditions where variation between pixels is negligible. The proposed MLBP uses a two stage encoding procedure which is more robust in detecting this variation in a local patch. A novel dimensionality reduction technique called Marginality Preserving Embedding (MPE) is also proposed for enhancing the face recognition accuracy. Unlike Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), which project data in a global sense, MPE seeks for a local structure in the manifold. This is similar to other subspace learning techniques but the difference with other manifold learning is that MPE preserves marginality in local reconstruction. Hence it provides better representation in low dimensional space and achieves lower error rates in face recognition. Two new concepts for robust face recognition are also presented in this dissertation. In the first approach, a neural network is used for training the system where input vectors are created by measuring distance from each input to its class mean. In the second approach, half-face symmetry is used, realizing the fact that the face images may contain various expressions such as open/close eye, open/close mouth etc., and classify the top half and bottom half separately and finally fuse the two results. By performing experiments on several standard face datasets, improved results were observed in all the new proposed methodologies. Research is progressing in developing a unified approach for the extraction of features suitable for accurate face recognition in a long range video sequence in complex environments.