Open Access

ARTICLE

BroadAttNet: Attention-Driven Micro-Expression Recognition

Hafiz Khizer bin Talib¹, Yanlong Cao², Muhammad Zaman^3,*, Sharifah Sakinah Syed Ahmad⁴, Nikola Ivkovic⁵, Mario Konecki⁵, Adnan Akhunzada⁶

1 State Key Laboratory of Fluid Power and Mechatronic Systems, School of Mechanical Engineering, Zhejiang University, Hangzhou, China
2 Key Laboratory of Advanced Manufacturing Technology of Zhejiang Province, School of Mechanical Engineering, Zhejiang University, Hangzhou, China
3 Department of Computer Science & IT, Superior University, Lahore, Punjab, Pakistan
4 Faculty of Artificial Intelligence and Cyber Security, Universiti Teknikal Malaysia Melaka, Durian Tunggal, Melaka, Malaysia
5 Faculty of Organization and Informatics, University of Zagreb, Pavlinska 2, Varazdin, Croatia
6 Department of Data and Cybersecurity, College of Computing & IT, University of Doha for Science and Technology, Doha, Qatar

* Corresponding Author: Muhammad Zaman. Email:

Computers, Materials & Continua 2026, 89(1), 55 https://doi.org/10.32604/cmc.2026.078779

Received 07 January 2026; Accepted 01 July 2026; Issue published 13 August 2026

Abstract

Micro-expression recognition (MER) is a demanding problem in affective computing because micro-expressions are brief, low-amplitude, involuntary facial movements that often reveal concealed affective states. Their recognition is complicated by weak muscle activation, short temporal duration, inter-subject variability, class imbalance, illumination changes, and the limited scale of publicly available MER datasets. To address these constraints, this paper introduces BroadAttNet, an attention-driven convolutional framework that embeds a Broadbent-inspired selective attention layer into a compact CNN backbone. The proposed layer learns to assign higher importance to discriminative facial regions while suppressing spatially redundant or noisy responses, thereby improving feature selectivity, interpretability, and recognition robustness without imposing substantial computational overhead. BroadAttNet is evaluated on three benchmark micro-expression datasets, SAMM, CASME II, and SMIC, using both K-Fold and Stratified K-Fold cross-validation protocols. The model achieves accuracies of 99.66% and 99.73% on SAMM, 98.73% and 97.65% on CASME II, and 98.62% and 97.09% on SMIC under K-Fold and Stratified K-Fold validation, respectively. Complementary ablation analysis, per-class evaluation, UAR reporting, model complexity analysis, and Grad-CAM visualization further demonstrate that the Broadbent Attention layer improves discriminative representation while preserving efficient inference. These findings indicate that BroadAttNet provides a robust, interpretable, and computationally practical solution for MER in human-computer interaction, behavioral analysis, and emotion-aware intelligent systems.

---

Keywords

---