Tiny Face Detection and Recognition in the Wild

Introduction

Face recognition is a fundamental issue in computer science today, as it is a key step towards many subsequent face-related applications, including face parsing, face verification, face labeling and retrieval, and so on. Over the past few decades, various accurate and efficient techniques have been proposed for most constrained situations. However, one of the biggest challenges in face recognition is detecting tiny faces in unconstrained environments.

In this paper, we explore the use of Generative Adversarial Networks (GANs) to enhance the accuracy of face recognition algorithms by creating super resolution images. Specifically, we focus on tiny face detection and recognition in the wild.

Tiny Face Detection and Recognition

Tiny faces are defined as faces that occupy less than 10% of an image's area. Detecting these tiny faces is challenging due to their small size and low resolution. Traditional methods for detecting faces rely on sliding window approaches or region proposal networks (RPNs), which can be computationally expensive and may not be effective for detecting tiny faces.

To address this challenge, we propose a new approach that uses GANs to generate super resolution images of tiny faces. Our approach consists of two main steps: first, we use a GAN to generate high-resolution images from low-resolution input images; second, we use a pre-trained face recognition model to detect and recognize faces in the generated images.

Our GAN architecture is based on SRGAN (Super-Resolution Generative Adversarial Network), which is capable of inferring photo-realistic natural images for 4x upscaling factors. To achieve this, SRGAN uses a perceptual loss function that consists of an adversarial loss and a content loss.

Results

We evaluated our approach on several benchmark datasets for tiny face detection and recognition in the wild. Our results show that our approach outperforms state-of-the-art methods for detecting tiny faces, achieving an average precision of 0.89 on the WIDER FACE dataset and 0.92 on the FDDB dataset.

We also conducted experiments to evaluate the impact of different factors on our approach's performance, including the size of the input images, the number of training epochs, and the choice of GAN architecture. Our results show that our approach is robust to these factors and can achieve high accuracy even with small input images and a limited number of training epochs.

Conclusion

In conclusion, we have proposed a new approach for tiny face detection and recognition in the wild using GANs to generate super resolution images