Facial Emotion Detection using CNN

The Facial Emotion Detection system is a deep-learning final-year project that finds faces in a live webcam feed and classifies the emotion on each face — happy, sad, angry, fear, surprise, disgust or neutral — using a convolutional neural network trained on the FER2013 dataset (35,887 grayscale 48×48 face images). Branded EmoSense, the Streamlit application overlays live emotion labels with confidence bars on the video feed and includes an analytics view showing the distribution of emotions across a session. With a trained model, training notebooks and real-time inference all included, this is one of the most viva-friendly deep-learning projects a student can present.

This is a built-to-order project: every deliverable is prepared fresh for you — full source code, trained model, report PDF, PPT and viva Q&A — so nothing looks recycled.

1. The webcam frame is captured and faces are detected with OpenCV's DNN face detector.
2. Each detected face is cropped, converted to grayscale, resized to 48×48 and normalized to match FER2013's input format.
3. The preprocessed face is passed through the trained CNN: stacked convolution + batch-normalization + max-pooling blocks, followed by dense layers with dropout and a 7-way softmax output.
4. The softmax probabilities become the emotion label (argmax) and the confidence bars shown in the UI.
5. Predictions are smoothed over a short sliding window to avoid flicker between adjacent emotions.
6. The analytics view aggregates per-session statistics — emotion distribution, dominant emotion, average confidence — rendered as charts.

• Real-time face detection with OpenCV DNN / Haar cascade, then CNN emotion classification
• CNN trained on FER2013: 35,887 images, 7 emotion classes, ~65–70% validation accuracy
• Live webcam inference at 15–25 FPS on CPU with per-face emotion labels
• Confidence bars for every prediction — shows model certainty, not just the top class
• EmoSense Streamlit app: live detection view + session analytics dashboard
• Emotion-over-time graph and distribution pie chart for a full demo session
• Snapshot capture: save frames with predicted emotions for the report
• Training notebook included — retrain, tune hyperparameters, plot learning curves
• Data augmentation pipeline (rotation, shift, zoom, flip) to reduce overfitting
• Multi-face support: detects and classifies several faces in one frame

• Complete, commented Python source code (training + inference + app)
• Pre-trained CNN model weights (.h5) — no GPU training needed on your machine
• Jupyter training notebook with augmentation, learning curves and confusion matrix
• Project report PDF (literature survey, CNN architecture, experiments, results, conclusion)
• PPT presentation for final review
• Viva Q&A preparation document (convolution math, FER2013 details, overfitting controls)
• Installation and run guide