Direct Hallucination: Direct Locality Preserving Projections (DLPP) for Face Super-Resolution

Faces captured by surveillance cameras are often of very low resolution. This significantly deteriorates face recognition performance. Super-resolution techniques have been proposed in the past to mitigate this. This paper proposes the novel use of a Locality Preserving Projections (LPP) algorithm called Direct Locality Preserving Projections (DLPP) for super resolution of facial images, or ldquoface hallucinationrdquo in other words. Because DLPP doesnpsilat require any dimensionality reduction preprocessing via Principle Component Analysis (PCA), it retains more discriminating power in its feature space than LPP. Combined with non-parametric regression using a generalized regression neural network (GRNN), the proposed work can render high-resolution face image from an image of resolution as low as 8x7 with a large zoom factor of 24. The resulting technique is powerful and efficient in synthesizing faces similar to ground-truth faces. Simulation results show superior results compared to other well-known schemes.

1. Introduction

Security needs of our increasingly intolerant world have increased manifold. Human face being a noninvasive biometric, is attractive for monitoring purposes. Video surveillance cameras abound in banks, airports, stores and parking lots, but in many cases, the face of interest in surveillance imagery is often small because of the distance between the camera and the object. The low image resolution of human face becomes a primary obstacle to face identification and recognition.