On Low-Resolution Face Recognition in the Wild: Comparisons and New Techniques (1805.11529v2)

Abstract: Although face recognition systems have achieved impressive performance in recent years, the low-resolution face recognition (LRFR) task remains challenging, especially when the LR faces are captured under non-ideal conditions, as is common in surveillance-based applications. Faces captured in such conditions are often contaminated by blur, nonuniform lighting, and nonfrontal face pose. In this paper, we analyze face recognition techniques using data captured under low-quality conditions in the wild. We provide a comprehensive analysis of experimental results for two of the most important applications in real surveillance applications, and demonstrate practical approaches to handle both cases that show promising performance. The following three contributions are made: {\em (i)} we conduct experiments to evaluate super-resolution methods for low-resolution face recognition; {\em (ii)} we study face re-identification on various public face datasets including real surveillance and low-resolution subsets of large-scale datasets, present a baseline result for several deep learning based approaches, and improve them by introducing a GAN pre-training approach and fully convolutional architecture; and {\em (iii)} we explore low-resolution face identification by employing a state-of-the-art supervised discriminative learning approach. Evaluations are conducted on challenging portions of the SCFace and UCCSface datasets.

• The paper demonstrates that sparse representation super-resolution (ScSR) consistently outperforms deep learning methods in enhancing severely degraded face images.
• The paper introduces center regularization to unify feature spaces between high- and low-resolution faces, achieving significant rank-1 improvements on benchmarks like SCface and UCCSface.
• The paper shows that leveraging DCGAN pre-training stabilizes training and reduces errors in face re-identification tasks across diverse datasets including MegaFace.

Low-Resolution Face Recognition Techniques: Insights and Comparisons

In the field of computer vision, face recognition has seen significant advancements, yet the Low-Resolution Face Recognition (LRFR) task remains a complex problem, especially under non-ideal conditions typical of surveillance scenarios. The paper "On Low-Resolution Face Recognition in the Wild: Comparisons and New Techniques" explores this niche domain by offering a comprehensive evaluation of face recognition methodologies for such degraded image scenarios.

Contributions of the Paper

The authors have delineated three prominent contributions:

1. Super-Resolution Techniques: The paper compares various super-resolution (SR) methods aimed at enhancing LR images before applying face recognition algorithms. Through extensive experimentation on datasets like AR and YouTube Faces, it is demonstrated that sparse representation super-resolution (ScSR) consistently outperforms modern deep learning techniques, particularly when handling extremely degraded images. This raises questions about the effectiveness of deep learning on lower quality datasets, which tend to introduce artifacts at lower resolutions.
2. Unified Feature Space and Discriminative Learning: By proposing center regularization, the authors introduce an innovative approach to bridge the gap between high-resolution (HR) and LR face images within a unified feature space. Evaluating on datasets such as SCface and UCCSface, the approach achieves significant rank-1 rate improvements, highlighting the efficacy of integrating strong intraclass clustering in deep feature learning.
3. Face Re-Identification and DCGAN Pre-training: The paper explores state-of-the-art face re-identification architectures and enhances them with fully convolutional layers and Spatial Pyramid Pooling (SPP). Moreover, leveraging a pre-trained DCGAN discriminator provides crucial initial weights for LR face images, facilitating transition across different datasets and reducing error rates. Training and testing on larger datasets like MegaFace revealed that pre-training on LR faces significantly stabilizes and improves training outcomes.

Practical Implications and Theoretical Insights

The findings present both practical and theoretical implications. Practically, these enhanced LRFR techniques could be pivotal in developing more reliable surveillance systems capable of recognizing individuals under challenging conditions. Super-resolution methods, particularly ScSR, could be harnessed to refine critical video feeds for effective monitoring. The theoretical underpinning regarding feature space unification informs future research on cross-resolution matching which is an area ripe for innovation.

The paper posits a substantial gap in recognizing real LR images against synthetic LR images derived from controlled HR sources. This puts further emphasis on collecting, managing, and utilizing naturally occurring LR datasets in research. Additionally, the DCGAN-based visualization and pre-training strategies open avenues for understanding LR dataset dynamics and address overfitting issues endemic to small datasets.

Outlook on Future Developments

Looking ahead, the proliferation of Generative Adversarial Networks (GANs) and their application in image processing presents exciting opportunities for LRFR. Advancements could focus on deepening our understanding of GAN-based models in recovering and correlating LR images to known HR personas. Investigations into unsupervised learning models could further diversify techniques for handling LR data. Furthermore, cross-disciplinary approaches integrating novel hardware, optimized data acquisition techniques, and AI could redefine the standards for LRFR.

This paper is a pertinent contribution to the LRFR field, advancing our understanding and highlighting areas for future digital surveillance technologies and methodologies.