Title: Research on multi-scale face detection based on graph embedded Swin-Transformer

Authors: Fang Wang; Huang Zhong

Addresses: School of Intelligent Manufacturing, Wuxi Vocational College of Science and Technology, Wuxi, Jiangsu, China ' Suzhou Triwin EST Co., Ltd., Suzhou, Jiangsu, China

Abstract: The existing Swin-Transformer has achieved great success in the field of object detection due to its lightweight, hierarchical and efficient long-distance dependency modelling advantages. However, in complex scenes, the impact of scale changes seriously restricts the performance of face detection. Therefore, this article proposes a multi-scale face detection network based on graph embedding Swin-Transformer. This network fully utilises multi-scale features and contextual information, effectively improving the performance of multi-scale face detection. Specifically, we designed a multi-feature fusion module based on Swin Transformer, which effectively integrates deep features of advanced information with shallow texture features. In addition, to effectively model the contextual information of the target, we use embedding graph convolution operations. We conducted extensive comparative experiments on publicly available data sets, and the experimental results showed that the proposed model achieved higher recognition performance.

Keywords: multi-scale face detection; Swin-Transformer; graph convolution embedded; multi-scale feature fusion.

DOI: 10.1504/IJCAT.2025.148169

International Journal of Computer Applications in Technology, 2025 Vol.76 No.1/2, pp.94 - 105

Received: 27 Aug 2023
Accepted: 14 Jun 2024
Published online: 27 Aug 2025 *