This project explores human emotion recognition from images using deep learning, focusing on CNN architectures including a baseline CNN, a modified VGGNet, and a pre-trained ResNet-50 with transfer learning. The FER-2013 dataset, containing grayscale facial expression images, was used for training and evaluation.
- Flask==3.0.0
- gunicorn==20.1.0
- tensorflow==2.8.0
- numpy==1.22.3
- Pillow==9.0.1
- matplotlib==3.5.1
- seaborn==0.11.2
- Werkzeug==3.0.0
FER-2013: Contains 35887 grayscale images of faces across seven emotion classes (0=Angry, 1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise, 6=Neutral)
- Baseline Model: Custom CNN
- VGGNet Modified CNN
- ResNet-50 with Transfer Learning
The models were evaluated based on accuracy, loss, and confusion matrices, with ResNet-50 showing superior performance (71% Test Accuracy)