This thesis presents a framework that integrates manifold learning and dimensionality reduction in 2-dimensional face recognition context. It aims to address the challenge of combined pose and illumination variations and the problem of limited samples in face database for face recognition.The integration of manifold learning and face recognition is presented as a concept of manifold learning autoencoder. An autoencoder consists of two parts, the encoder and the decoder. The first part of the thesis focuses on the encoder, where manifold learning algorithms learn and extract the variability in facial images, particularly pose and illumination variations, and project them onto a lower dimensional space. In the low-dimensional space, these variations form a complicated manifold structure that may be intersected or overlapped with each other. The capabilities of linear and nonlinear methods to untangle these variations are quantitatively and qualitatively compared in this thesis. The findings for this analysis show that none of the single manifold methods, linear or nonlinear, can sufficiently represent the combined pose and illumination manifold. This justifies the need of a multi-manifold method that allows each manifold to be represented accurately.The second part of this thesis introduces an implementation of the decoder, where high-dimensional data (images) can be reconstructed based on the learned low-dimensional projection. This concept is translated into face relighting and pose synthesis processes, where an image is augmented to contain novel pose or illumination condition that was extracted by manifold learning. By synthesising images with novel variations, it leverages the capability of face recognition algorithm by minimising the difference between the sample image and a test image that can contain arbitrary pose and illumination variations. Finally, the face relighting and pose synthesis processes are combined into a multi-manifold face recognition framework (MMFR) to handle facial images with combined pose and illumination variations. A series of face recognition experiments were conducted to validate the MMFR and have shown encouraging recognition results on a benchmark face database.