Discrete Orthonormal Stockwell Transform Based Feature Extraction for Pose Invariant Face Recognition

The statistical description of the face varies drastically with changes in pose. These variations make Face Recognition (FR) even more challenging. In this paper, four novel techniques are proposed, viz., Discrete Orthonormal Stockwell Transform (DOST), Entropy based image resizing, Denoising and Enhanced Training and Testing methodology, to improve the performance of an FR system. DOST is used for efficient Feature Extraction and Entropy based image resizing calculates the Optimum Resizing Factor (ORF) to be used. Denoising is done through a combination of Deghosting, Intensity Mapped Unsharp Masking and Anisotropic Diffusion. Pose neutralization is achieved through the enhanced Training and Testing methodology. Thus, a complete FR system for enhanced recognition performance is presented. Experimental results on four benchmark face databases, namely, CAS-PEAL R1, Color FERET, FEI and HP, illustrate the promising performance of the proposed techniques for face recognition.