← 소식
[논문게재] PATTERN RECOGNITION
- DATE
- 2025.09.03
- TYPE
- Academic
저자: Kyeong Tae Kim and Jae Young Choi
출판지: PATTERN RECOGNITION
논문제목: Triple-branch Attention Network with Multiscale Facial Region Feature Selection for Facial Emotion Recognition
Abstract
This paper proposes new and novel “Triple-branch attention network (TriBAN)”, which selects and integrates key attention features based on multiple-scale facial regions for improved facial expression recognition (FER). The proposed TriBAN consists of three main branches: global branch, large local branch, and small local branch. The aim of global branch is to capture the overall characteristics of facial image. In addition, the large local branch is equipped with attention components within each regional area to filter out irrelevant regions, addressing occlusion and pose variations, and simultaneously capturing discriminant attention features, while the small local branch enhances the robustness of edge information and upscales feature size through pixel shuffling, enabling it to capture subtle expression changes in the generated small regions. In each aforementioned branch, discriminative features are adaptively chosen, and graph convolutional network is effectively used for integrating the selected features by considering their structural relationships. Extensive and comparative experiments on three wild FER datasets, namely RAF-DB, AffectNet, and FERplus, showed that the proposed TriBAN achieves better or comparable FER performances, compared to other recently developed FER methods. Our method outperformed FER approaches using ResNet18 as the backbone and showed slightly lower performance than ViT-based algorithms but achieved more than twice the inference speed.
출판지: PATTERN RECOGNITION
논문제목: Triple-branch Attention Network with Multiscale Facial Region Feature Selection for Facial Emotion Recognition
Abstract
This paper proposes new and novel “Triple-branch attention network (TriBAN)”, which selects and integrates key attention features based on multiple-scale facial regions for improved facial expression recognition (FER). The proposed TriBAN consists of three main branches: global branch, large local branch, and small local branch. The aim of global branch is to capture the overall characteristics of facial image. In addition, the large local branch is equipped with attention components within each regional area to filter out irrelevant regions, addressing occlusion and pose variations, and simultaneously capturing discriminant attention features, while the small local branch enhances the robustness of edge information and upscales feature size through pixel shuffling, enabling it to capture subtle expression changes in the generated small regions. In each aforementioned branch, discriminative features are adaptively chosen, and graph convolutional network is effectively used for integrating the selected features by considering their structural relationships. Extensive and comparative experiments on three wild FER datasets, namely RAF-DB, AffectNet, and FERplus, showed that the proposed TriBAN achieves better or comparable FER performances, compared to other recently developed FER methods. Our method outperformed FER approaches using ResNet18 as the backbone and showed slightly lower performance than ViT-based algorithms but achieved more than twice the inference speed.