Very Low-Resolution Face Recognition In The Wild

Abstract

Face recognition has found extensive application scenarios in the United States military. However, the face region from real photos or videos could be very small and low-quality, which constitutes the very low-resolution face recognition and fails most existing methods. The objective of this project is to create new methods of joint super-resolution and recognition for face recognition from very low resolution face images, which will lead to creating an optimized pipeline for very low resolution face recognition in the wild. The success of this research will lay the groundwork for automated and robust object recognition under poor data quality, and contribute timely to the accomplishment of the Army missions. The crux of the proposed approach spans a three-layer hierarchy: I) On data level, research will improve the resolution of very low resolution faces to a reasonable scale, by utilizing joint examples and adopting a task-driven loss, with multi-frame references when available; 2) On model level, research will synthesize models obtained from multiple resolution scales to boost the overall recognition performances, and learn optimal metrics for each scale to keep the affinity structures identical; 3) On system level, research will jointly optimize the face alignment step with the SR and recognition modules, as a closed-loop system.

Document Details

Document Type
DoD Grant Award
Publication Date
Jan 12, 2017
Source ID
W911NF1510317

Entities

People

  • Thomas Huang

Organizations

  • Army Contracting Command
  • United States Army
  • University of Illinois Urbana–Champaign

Tags

Readers

  • Computer Vision.
  • Distributed Systems and Data Platform Development
  • Strategic Security Studies

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks