This project implements a Convolutional Neural Network (CNN) for facial emotion recognition using the FER-2013 dataset.
- Grayscale image preprocessing (48×48)
- CNN built from scratch
- Data augmentation
- Batch normalization & dropout
- Early stopping to prevent overfitting
FER-2013
Validation accuracy stabilizes around ~45–55% due to the noisy nature of the dataset and due to similarity of human face in different emotions. Tested model on a random internet photo and works fine.
- Having more data
- Regularisation can be done upto some extend
- Transfer learning (MobileNet / EfficientNet)
- Face detection before classification
Rishi57