Use this URL to cite or link to this record in EThOS:
Title: Face Recognition Based on Histogram And Spin Image
Author: Li, Yang
Awarding Body: University of York
Current Institution: University of York
Date of Award: 2007
Availability of Full Text:
Access through EThOS:
This thesis presents our research work on shape-based human face recognition exploiting curvature-based histogram and spin image. Instead of the popular 2D shape information represented by fiducial points, the novelty here is the use of 2.5D shape information obtained by shape-from-shading (SFS). Though surface normals generated by performing shape-from-shading on objects are not widely accepted as a precise shape representation for face recognition purposes, recent research in shape-from-shading [Ragheb and Hancock, 2003] [prados et aI., 2006] [Castelan, 2006] [Smith and Hancock, 2006] has made it possible to recover fairly accurate shape information under various conditions from face images in the real world. These contributions make our face recognition approaches based on 2.5D shape recovered from a single image possi.ble. Chapter 2 is a thorough review of the existing literature in the areas pf surface reconstruction using shape-from-shading, appearance-based and model-based face recognition on 2D and 3D data, and histogram-based image representation and recognition. The literature is pretty sparse on works on shape-from-shading in the face recognition area, however there are plenty of approaches based on 2D images ~ and 3D range data. With accurate height map being recovered from single face image [Prados et aI., 2006] [Castelan, 2006] and statistical model being proposed to recover surface normals [Smith and Hancock, 2006], we have enough 2.5D shape infonnation recovered' from 2D image based on which we can perfonn face recognition. In Chapter 3, we present our curvature-based histogram appro,ach as our first contribution, which employs principal curvature infonnation calculated from the Hessian matrix based on the recovered surface nonnals. Generalized entropies are introduced as similarity measurements, which give stable performance especially when the number of relative images varies. While curvature-based histogram proves to be a fairly good face recognition approach with concise representation and easy comparison, its performance is not always stable when applied to different databases. Therefore more advanced facerecognition approach is required to bring better and more stable identification results. In Chapter 4, we derive patch-based spin image as a local shape representation from the idea of global curvature-based histogram in Chapter 3. This representation is inspired by [Johnson and Hebert, 1999] and adapted in this thesis as a solution to face recognition problem. Instead of using 3D range data, the estimated height map reconstructed by shape-from-shading is employed in the spin image construction. Also the mean needle map model is used as the preprocessing to correct the errors and noise that exists in the surface nonnal estimates. Face surface is segmented into small patches and the spin image corresponding to the surface is composed of histograms constructed on each surface patch. In Appendix A, we propose dual spin image to address the difficulties of recognizing face in rotation, among which the two major problems of point correspondence and point occlusion are of particular importance. Therefore we propose the idea of neighbour area spin image to construct the pointwise neighbourhood surface feature collection and use the ratio of projected distances to relative angles to alleviate the errors introduced by surface rotation. We also present the linear model and the finite Gaussian mixture model to approximate novel dual spin image using the existing dual spin images. Face recognition is performed based on the model parameters. .The work in this thesis suggests that 2.5D shape feature recovered by shapefrom- shading can be used for the purpose offace recognition. Also the appearancebased approaches derived from histogram are effective for face recognition. The result suggests that shape information recovered from single image is sufficient for face recognition based on the condition that shape-from:-shading can successfully recover the surface normal fields and the height map from the image.
Supervisor: Not available Sponsor: Not available
Qualification Name: University of York, 2007 Qualification Level: Doctoral
EThOS ID:  DOI: Not available