Enhancing Multi-Orientation Face Detection Using Receptive Field Block Modules in the RetinaFace Architecture

Mohammad Abdan Syakura, Wahyono, S.Kom., Ph.D.

2025 | Skripsi | ILMU KOMPUTER

Multi-orientation facial detection is one of the most challenging tasks in computer vision. In this research, a RetinaFace-based face detection model is proposed, integrating the Receptive Field Block (RFB) module into the Feature Pyramid Network (FPN) layers to improve the detection of multi-scale and multi-orientation features. The model employs a MobileNetV2 backbone for computational efficiency. To assess the effectiveness of the proposed modifications, its performance was compared against a baseline RetinaFace model through evaluations on WIDER FACE, AFLW2000 (at varied resolutions), and HPID (across diverse poses and resolutions). The results consistently show that the RFB-enhanced model outperforms the baseline, demonstrating improvements in detection recall and overall performance, especially under challenging conditions such as low image resolution and extreme head poses where the baseline model’s effectiveness diminishes. This work contributes to the development of face detection systems better able to accurately identify faces across a wider range of scales and orientations.

Multi-orientation facial detection is one of the most challenging tasks in computer vision. In this research, a RetinaFace-based face detection model is proposed, integrating the Receptive Field Block (RFB) module into the Feature Pyramid Network (FPN) layers to improve the detection of multi-scale and multi-orientation features. The model employs a MobileNetV2 backbone for computational efficiency. To assess the effectiveness of the proposed modifications, its performance was compared against a baseline RetinaFace model through evaluations on WIDER FACE, AFLW2000 (at varied resolutions), and HPID (across diverse poses and resolutions). The results consistently show that the RFB-enhanced model outperforms the baseline, demonstrating improvements in detection recall and overall performance, especially under challenging conditions such as low image resolution and extreme head poses where the baseline model’s effectiveness diminishes. This work contributes to the development of face detection systems better able to accurately identify faces across a wider range of scales and orientations.

Kata Kunci : face detection, multi-orientation, multi-scale, deep learning, retinaface, receptive field block, mobilenetv2