BinaryFaceNet: A Binarized Approach for Real-Time Very Low Resolution Face Recognition in Video Surveillance Scenarios - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2024

BinaryFaceNet: A Binarized Approach for Real-Time Very Low Resolution Face Recognition in Video Surveillance Scenarios

Résumé

With the rise of Deep Learning methods in the past decade, we have seen a tremendous leap in accuracy performance for various computer vision tasks, such as face recognition. Training and inference for these methods often require expensive hardware equipment. In more recent years, efficient Deep Learning based methods have emerged, with Lightweight and Quantized Convolutional Neural Networks being useful in computer vision applications for constrained hardware platforms. However, most approaches have been proposed for general-purpose tasks in mind. In this paper we propose BinaryFaceNet, the first approach specific for very low resolution face recognition in the wild using an original Quantized Convolutional Neural Network design. We propose to reduce the overhead of operations when down-sizing feature maps with grouped convolutions, using a progressive downsizing in the Squeeze-Excite block for a more accurate weighted signal and novel Global Feature Extraction and Local Feature Extraction blocks with DepthWise convolutions and alternated residual connections to avoid signal degradation due to quantization and network depth. With our implementation, we reduce Multiply-Accumulate operations in half with respect to the state of the art in Quantized Convolutional Neural Networks while achieving a limited compromise in recognition performance in Very Low Resolution benchmark datasets.
Fichier non déposé

Dates et versions

hal-04393666 , version 1 (14-01-2024)

Licence

Paternité

Identifiants

  • HAL Id : hal-04393666 , version 1

Citer

Luis S. Luevano, Leonardo Chang, Miguel González-Mendoza, Yoanna Martínez-Díaz, Heydi Méndez-Vázquez, et al.. BinaryFaceNet: A Binarized Approach for Real-Time Very Low Resolution Face Recognition in Video Surveillance Scenarios. 2024. ⟨hal-04393666⟩
31 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More