Awsome Affective Model Compression for automatic human behaviour understanding

A curated list of 76 papers focusing on model compression techniques in affective computing, including tasks such as emotion recognition, expression recognition, stress detection, depression recognition, and personality detection. Categorized by task, modality, and compression method.

Emotion recognition

👁️ Visual Modality

• Light-FER: A Lightweight Facial Emotion Recognition System on Edge Devices
  Method: Pruning and Quantization
  Dataset: FER2013
  Link

• Factorized Higher-Order CNNs with an Application to Spatio-Temporal Emotion Estimation
  Method: CP Tensor Decomposition
  Dataset: SEWA
  Link

• A Lightweight Method for Face Expression Recognition Based on Improved MobileNetV3
  Method: Lightweight (Improved MobileNetV3)
  Datasets: FER2013, RAF-DB
  Link

• Three Convolutional Neural Network Models for Facial Expression Recognition in the Wild
  Method: Lightweight (Depthwise Separable CNN)
  Datasets: FER2013, RAF-DB
  Link

• Comparison of Different Depth of Convolutional Neural Network Models for Facial Expression Recognition
  Method: Lightweight (Shallow CNN)
  Dataset: FER2013
  Link

• Facial Emotion Recognition Through Custom Lightweight CNN Model: Performance Evaluation in Public Datasets
  Method: Lightweight (Custom Low-Parameter CNN)
  Datasets: FER2013, RAF-DB, AffectNet, CK+
  Link

• MobileNetV2: Inverted Residuals and Linear Bottlenecks
  Method: Lightweight (Inverted Bottleneck Convolution)
  Note: Foundational for later work using MobileNetV2 structure
  Link

• Pseudo-Inverted Bottleneck Convolution for DARTS Search Space
  Method: Lightweight (Inverted Bottleneck Convolution)
  Datasets: RAF-DB, FER2013H
  Link

🔊 Audio Modality

• Discriminative pruning of deep neural networks for speech emotion recognition
  Method: Pruning
  Dataset: IEMOCAP
  Link

• Knowledge distillation using HuBERT embeddings for small-footprint emotion recognition
  Method: Knowledge Distillation
  Dataset: IEMOCAP
  Link

• LIGHT-SERNET: A Lightweight Deep Learning Architecture for Speech Emotion Recognition
  Method: Lightweight
  Dataset: RAVDESS
  Link

• Parallel attention-based CNN model for speech emotion recognition
  Method: Lightweight
  Dataset: IEMOCAP
  Link

• Low-footprint convolutional model for real-time speech emotion recognition
  Method: Lightweight
  Dataset: EmoDB
  Link

• Structured pruning for lightweight SER on mobile devices
  Method: Pruning
  Dataset: RAVDESS
  Link

• Multiscale CNN with quantization-aware training for efficient SER
  Method: Quantization
  Dataset: IEMOCAP
  Link

• Compact transformers for end-to-end speech emotion recognition
  Method: Lightweight
  Dataset: IEMOCAP
  Link

• Speech emotion recognition with tiny speech CNNs
  Method: Lightweight
  Dataset: EmoDB
  Link

• Efficient CNN architecture using spectrogram compression for SER
  Method: Compression (custom)
  Dataset: RAVDESS
  Link

🧠 Physiological Modality

• Disentangling EEG Representation Using Neuroscience Priors for Emotion Recognition
  Method: Pruning
  Dataset: DEAP
  Link

• A Teacher–Student Framework for Emotion Recognition Using EEG Signals
  Method: Knowledge Distillation
  Dataset: DEAP
  Link

• A Quantized CNN for Emotion Recognition from EEG Signals
  Method: Quantization
  Dataset: DEAP
  Link

• Temporal Convolutional 3D Network for Emotion Recognition with EEG
  Method: Lightweight
  Dataset: SEED
  Link

• SHAP-guided Pruning of GCNs for EEG-based Emotion Recognition
  Method: Pruning
  Dataset: AMIGOS
  Link

• Quantization-aware Training for LSTM in Emotion Recognition using PPG
  Method: Quantization
  Dataset: WESAD
  Link

• Tiny 1D CNN for Real-time ECG-based Emotion Classification
  Method: Lightweight
  Dataset: DRIVE
  Link

• Distilled EEGNet for Real-time Affective State Detection
  Method: Knowledge Distillation
  Dataset: DEAP
  Link

• Hybrid CNN-LSTM with Pruned Structure for Emotion Detection
  Method: Pruning
  Dataset: DEAP
  Link

• Low-complexity Temporal CNN for Wearable EEG Emotion Recognition
  Method: Lightweight
  Dataset: DREAMER
  Link

• EEG Emotion Recognition via Quantized Mobile CNN
  Method: Quantization
  Dataset: MAHNOB-HCI
  Link

🔄 Multimodal Modality

• Efficient Audio-Visual Emotion Recognition Using Structured Pruning
  Method: Pruning
  Dataset: MuSe-Humor
  Link

• Cross-Modal Knowledge Distillation for Multimodal Emotion Recognition
  Method: Knowledge Distillation
  Dataset: RECOLA
  Link

• Tensor Fusion Network Compression via Tensor Decomposition
  Method: Decomposition
  Dataset: CMU-MOSEI
  Link

• Mobile Multimodal Emotion Recognition Using Lightweight Fusion Network
  Method: Lightweight
  Dataset: IEMOCAP
  Link

• Distilling Emotion Representations from Audio-Visual Models to Unimodal Models
  Method: Knowledge Distillation
  Dataset: CMU-MOSEI
  Link

• A Low-Resource Multimodal Transformer for Emotion Recognition
  Method: Lightweight
  Dataset: CMU-MOSEI
  Link

• Multimodal Emotion Recognition with Compact Multiscale Attention Fusion
  Method: Lightweight
  Dataset: IEMOCAP
  Link

• Multimodal Emotion Recognition with Feature-level Compression and KD
  Method: Knowledge Distillation + Feature Compression
  Dataset: CMU-MOSEI
  Link

Expression Recognition

👁️ Visual Modality

• Magnitude-based Pruning for Facial Expression Recognition on Mobile Devices
  Method: Pruning
  Dataset: RAF-DB
  Link

• Facial Expression Recognition Based on Pruning Optimization Technology
  Method: Pruning
  Dataset: CK+, JAFFE
  Link

• Adaptive CNN Pruning Based on Structural Similarity of Filters (APSSF)
  Method: Pruning
  Dataset: FER2013
  Link

• Teacher-Bounded Loss for FER Knowledge Distillation
  Method: Knowledge Distillation
  Dataset: AffectNet
  Link

• Multiple Specialized Teachers Knowledge Distillation for Fair Face Recognition
  Method: Knowledge Distillation
  Dataset: FairFace
  Link

• Distilled VGG16 for Real-time FER
  Method: Knowledge Distillation
  Dataset: FER2013
  Link

• FER using MobileNet with Distilled Attention
  Method: Knowledge Distillation
  Dataset: FER2013
  Link

• Distilled ResNet with Emotion Prior Masks for FER
  Method: Knowledge Distillation
  Dataset: FERPlus
  Link

• IRBN: Iterative Residual Binarized Network for Efficient FER
  Method: Quantization
  Dataset: RaFD
  Link

• BitNet: Binary CNN for Embedded FER
  Method: Quantization
  Dataset: AffectNet
  Link

• Weight Quantization of CNNs for Expression Recognition
  Method: Quantization
  Dataset: JAFFE
  Link

• Ternary Quantization-aware FER CNN
  Method: Quantization
  Dataset: RAF-DB
  Link

• BinaryNet for Facial Expression Recognition on Edge Devices
  Method: Quantization
  Dataset: FER2013
  Link

• Quantized Residual Networks for FER in Unconstrained Environments
  Method: Quantization
  Dataset: FERPlus
  Link

• Low-rank Approximation for Expression Recognition Networks
  Method: Decomposition
  Dataset: FER2013
  Link

• EfficientNet-lite for Mobile Facial Expression Recognition
  Method: Lightweight
  Dataset: AffectNet
  Link

• Attention-based Shallow CNN for FER with Fusion Layer
  Method: Lightweight
  Dataset: RAF-DB
  Link

• MobileFaceNet-based FER with Tiny Parameter Count
  Method: Lightweight
  Dataset: FER2013
  Link

• Compact Residual Attention Network for FER
  Method: Lightweight
  Dataset: RAF-DB
  Link

• FER with Lightweight Transformer and Attention
  Method: Lightweight
  Dataset: AffectNet
  Link

• AutoFER: Neural Architecture Search for Facial Expression Recognition
  Method: Lightweight
  Dataset: CK+
  Link

• Tiny Dual-branch CNN for FER in Video
  Method: Lightweight
  Dataset: AFEW
  Link

• Efficient Deep Feature Compression for FER on the Edge
  Method: Compression (custom)
  Dataset: AffectNet
  Link

• Multi-scale Lightweight CNN for FER in Real-World Scenarios
  Method: Lightweight
  Dataset: AffectNet
  Link

• Compact CNN using Group Convolutions for FER
  Method: Lightweight
  Dataset: RAF-DB
  Link

• Lightweight Attention-Guided FER with Mobile Efficiency
  Method: Lightweight
  Dataset: AffectNet
  Link

• Pruned ResNet-18 for FER with Knowledge Guidance
  Method: Pruning + Knowledge Distillation
  Dataset: RAF-DB
  Link

Stress Detection

📝 Text Modality

• Compact BERT for Stress Detection from Social Media Posts
  Method: Lightweight
  Dataset: Dreaddit
  Link

🧠 Physiological Modality

• Real-time Stress Detection using Quantized LSTM on Wearable PPG
  Method: Quantization
  Dataset: WESAD
  Link

• TinyCNN-STRESS: Lightweight CNN for Stress Detection from ECG
  Method: Lightweight
  Dataset: DRIVE
  Link

• Pruned Deep CNN for EEG-based Stress Detection
  Method: Pruning
  Dataset: DEAP
  Link

• Distilled BiLSTM for Wearable EEG-based Stress Monitoring
  Method: Knowledge Distillation
  Dataset: WESAD
  Link

Depression Detection

👁️ Visual Modality

• Efficient Deep Learning Framework for Visual Depression Detection Using Micro-expressions
  Method: Lightweight
  Dataset: DAIC-WOZ
  Link

• Distilled Video-based Model for Depression Severity Assessment
  Method: Knowledge Distillation
  Dataset: AVEC
  Link

🧠 Physiological Modality

• Quantized LSTM for Wearable EEG Depression Detection
  Method: Quantization
  Dataset: WESAD
  Link

• Lightweight CNN with HRV features for Depression Classification
  Method: Lightweight
  Dataset: DREAMER
  Link

• Pruned EEGNet for Depression Risk Stratification
  Method: Pruning
  Dataset: DEAP
  Link

Personality Assessment

📝 Text Modality

• Distilled BERT Model for Efficient Personality Trait Detection
  Method: Knowledge Distillation
  Dataset: Essays Dataset
  Link