Deep Learning Project : Facial Emotion Recognition with KAN

Introduction

This project focuses on enhancing facial emotion recognition through a novel deep learning approach. We aim to leverage the Kolmogorov-Arnold Network (KAN) to improve the accuracy and interpretability of emotion classification.

Dataset

The FER2013 dataset is utilized, comprising 35,887 grayscale images categorized into seven emotions: Angry, Disgust, Fear, Happy, Sad, Surprise, and Neutral.

Methodology

Our approach builds upon a Convolutional Neural Network (CNN) architecture, similar to the baseline established by GSNCodes. The key innovation lies in replacing traditional dense layers with a Kolmogorov-Arnold Network (KAN) to potentially enhance non-linear function approximation capabilities.

Evaluation Metrics

Performance will be assessed using standard metrics:

• Accuracy
• Precision
• Recall
• F1-score

Results

Final Test Accuracy: 69.21%

Baseline

https://github.com/GSNCodes/Emotion-Detection-FER2013/blob/master/Emotion_Recognition_Train.ipynb Baseline Final Test Accuracy: 63%

Detailed Classification Report:

              precision    recall  f1-score   support

       Angry       0.59      0.61      0.60       491
     Disgust       0.77      0.60      0.67        55
        Fear       0.56      0.56      0.56       528
       Happy       0.88      0.90      0.89       879
         Sad       0.61      0.51      0.56       594
    Surprise       0.84      0.77      0.80       416
     Neutral       0.61      0.70      0.65       626

    accuracy                           0.69      3589
   macro avg       0.69      0.67      0.68      3589
weighted avg       0.69      0.69      0.69      3589

Conclusion

We anticipate the CNN + KAN integrated model to outperform the baseline CNN in accuracy and other metrics, offering improved recognition of facial emotions and potentially greater model interpretability.