Article: Rotation-invariant face detection with guided deformable attention Journal: International Journal of Information and Communication Technology (IJICT) 2024 Vol.25 No.8 pp.32 - 48 Abstract: Detecting rotated faces has always been a challenging task. Fixed convolutional kernels struggle to effectively match features after rotation, while the sampling point offsets of deformable convolutions are limited by complex backgrounds. To address this issue, we propose a guided deformable attention (GDA) network. Guiding the offset direction of sampling points by adding constraints of facial structure to deformable convolutions. The GDA network adopts a dual-stream structure, with one branch detecting the inherent structural information for preliminary positioning of the face area; then, the second branch uses deformable convolution to perform pixel-level feature extraction on the face within the range. In addition, we introduce a novel loss, which, during the guidance process, aligns the activation areas in the feature maps extracted by the two branches through the KL divergence. Extensive experimental results validate that GDA network performs excellently on multiple face detection datasets, surpassing the current state-of-the-art face detection methods. Inderscience Publishers - linking academia, business and industry through research

Title: Rotation-invariant face detection with guided deformable attention

Authors: Bin Deng; Guanghui Deng

Addresses: College of Computer Science, Hunan University of Technology, Zhuzhou, Hunan – 412007, China ' College of Science, Hunan University of Technology, Zhuzhou, Hunan – 412007, China

Abstract: Detecting rotated faces has always been a challenging task. Fixed convolutional kernels struggle to effectively match features after rotation, while the sampling point offsets of deformable convolutions are limited by complex backgrounds. To address this issue, we propose a guided deformable attention (GDA) network. Guiding the offset direction of sampling points by adding constraints of facial structure to deformable convolutions. The GDA network adopts a dual-stream structure, with one branch detecting the inherent structural information for preliminary positioning of the face area; then, the second branch uses deformable convolution to perform pixel-level feature extraction on the face within the range. In addition, we introduce a novel loss, which, during the guidance process, aligns the activation areas in the feature maps extracted by the two branches through the KL divergence. Extensive experimental results validate that GDA network performs excellently on multiple face detection datasets, surpassing the current state-of-the-art face detection methods.

Keywords: rotated face detection; deformable convolution; attention; KL divergence; dual-stream; pixel-level.

DOI: 10.1504/IJICT.2024.142299

International Journal of Information and Communication Technology, 2024 Vol.25 No.8, pp.32 - 48

Received: 24 Jul 2024
Accepted: 09 Sep 2024
Published online: 17 Oct 2024 *

Title: Rotation-invariant face detection with guided deformable attention

Keep up-to-date