Recognizing Rotated Faces from Two Orthogonal Views in Mugshot Databases

Abstract

Tolerance to pose variations is one of the key remaining problems in face recognition. It is of great interest in airport surveillance systems using mugshot databases to screen travellers' faces. This paper presents a novel pose-invariant face recognition approach using two orthogonal face images from mugshot databases. Virtual views under different poses are generated in two steps: shape modeling and texture synthesis. In the shape modeling step, a feature-based multilevel quadratic variation minimization approach is applied to generate smooth 3D face shapes. In the texture synthesis step, a non-Lambertian reflectance model is explored to synthesize facial textures taking into account both diffuse and specular reflections. A view-based face recognizer is used to examine the feasibility and effectiveness of the proposed pose-invariant face recognition. The experimental results show that the proposed method provides a new solution to the problem of recognizing rotated faces