Facial expression recognition through multi-level features extraction and fusion

Xie, Yuanlun; Tian, Wenhong; Zhang, Hengxin; Ma, Tingsong

doi:10.1007/s00500-023-08531-z

Cited by 8 publications

(2 citation statements)

References 39 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Yu et al [16] introduced a feature fusion strategy to fuse local texture features of the face, face edge detection features, and the face image, so as to obtain more complete image features and more in-depth high-level semantic features. Xie et al [17] proposed the use of a multi-scale convolutional kernel to extract facial expression features at different semantic levels. In addition, they used both local and global attention to fuse image features at different levels in order to enhance the representation of facial expression features.…”

Section: Related Work 21 Deep Learning In Fermentioning

confidence: 99%

“…During training process, we applied common data August 10,2023 enhancement strategies including random level flipping, random cropping and rotation. We used Overall(8 cls) MFEF [17] 0.8808 -0.8811 -0.5938 CT-DBN [19] 0.8840 -0.8917 --Face2Exp [43] 0.8854 --0.6423 -EAC [42] 0.8999 -0.8964 0.6532 -MPCSAN [7] 0.9016 -0.8991 -0.6158 ARM [44] 0.9042 0.8277 -0.6520 0.6133 PIDVIT [18] 0…”

Section: Implementation Detailsmentioning

confidence: 99%

See 1 more Smart Citation

Facial Expression Recognition with Enhanced Relation-Aware Attention and Cross-Feature Fusion transformer

DONG,

Wang,

et al. 2024

Preprint

View full text Add to dashboard Cite

Face expression recognition(FER) is an important research branch in the field of the computer vision neighborhood. Three prevalent problems in FER tasks that severely impact recognition rates are inter-class similarity, intra-class differences, and facial occlusion issues. Although there have been studies that address some of these issues, none of them can adequately address all three issues in a unified framework. In this paper, we propose a novel dual-branch structure of enhanced relation-aware attention and cross-feature fusion transformer network to comprehensively solve all three issues. Specifically, we design the Enhanced Relation-Aware Attention module to maximize the exploration of more local expression features. At the same time, the Transformer Perceptual Encoder module is adopted to establishing the contextual relationship between individual patches under global information. This greatly alleviates the inter-class similarity problem and the facial occlusion and facial pose transformation problems. On the basis of a dual branch structure, we extract facial image features using facial landmarks features to guide them and design Cross-Feature Fusion Transformer module to deeply cross-fuse two different semantic features. Experiments are performed and results show that our method can greatly alleviated intra-class difference problem with comparison of several traditional methods on three commonly used datasets.

show abstract

Section: Related Work 21 Deep Learning In Fermentioning

confidence: 99%