HadaSmileNet

Hadamard fusion of handcrafted and deep-learning features for genuine smile recognition (IEEE ICDM 2025)

IEEE ICDM 2025 (Oral + Poster)

HadaSmileNet introduces a novel Hadamard-fusion framework that integrates handcrafted facial features with deep learning representations for enhanced genuine smile recognition. This work was accepted for both oral and poster presentation at IEEE ICDM 2025.

Key Achievements

  • 26% Efficiency Gain: Reduced parameters, training time, and inference latency by 26%
  • Superior Accuracy: Outperformed multi-task learning baselines while maintaining accuracy
  • Hadamard Fusion: Novel integration method for handcrafted and learned features
  • Clinical Relevance: Collaborated with domain experts to ensure real-world applicability

Technical Approach

Architecture: Hadamard product-based fusion of handcrafted Duchenne Marker features and transformer-learned features

Datasets: Validated on 4 benchmark smile recognition datasets

Framework: PyTorch, Transformers

Efficiency: 26% reduction in computational requirements vs. multi-task baselines

Impact

This research advances facial emotion recognition with an efficient, interpretable approach that combines domain knowledge (Duchenne Marker) with deep learning, making it suitable for resource-constrained applications in healthcare and HCI.

Status: Published at IEEE ICDM 2025 (November 15, 2025) Authors: Mohammad Junayed Hasan, Nabeel Mohammed, Shafin Rahman, Philipp Koehn

References