Facial Expression Recognition

A facial expression classifier: a CNN trained from scratch on 48×48 grayscale face crops, evaluated on a held-out test set and served through a FastAPI web app that reads from the webcam.

View code on GitHub Run it locally

65%test accuracy · 7 classes

01 · Problem

What needed solving

Faces at 48×48 pixels carry little detail and the classes are very uneven: the test set has 111 'disgust' images and 1,774 'happy' ones. A single accuracy number hides how the model does on the rare classes.

02 · Solution

How it works

A mid-sized CNN (Conv2D blocks with BatchNormalization, MaxPooling, Dropout and GlobalAveragePooling), a validation split taken automatically from the training data, a final evaluation only on the untouched test set, and one preprocessing path shared by training, evaluation and live prediction.

03 · Key features

What it does

Face detection with OpenCV's Haar cascade, then crop, grayscale and resize to 48×48
65.0% test accuracy over 7 classes; F1 0.87 for happy and 0.77 for surprise
Training curves, per-class report and confusion matrix saved as artefacts
Single-image prediction script and a FastAPI web app with live webcam input

04 · Tech stack

Built with

Python
Keras
OpenCV
FastAPI

Status: In progress
Code: github.com/KacperBlok1/face-emotion-recognition-cnn
Demo: Not hosted yet · run locally

05 · Challenges

What was tricky

Rare classes
Recall is 22% for 'disgust' and 30% for 'fear': with so few examples the model learns them poorly. The per-class report and confusion matrix make that visible instead of hiding it behind the overall 65%.
Same preprocessing everywhere
If the webcam app prepared faces differently from training, accuracy would drop without any error. One shared preprocessing function is used for training, evaluation and prediction.
Dependency versions
TensorFlow 2.10 needs numpy<2 and protobuf<3.20, so versions are pinned and OpenCV is installed with --no-deps to stop pip from upgrading numpy.

06 · Screenshots

How it looks

Training and validation loss and accuracy per epoch — Training history: loss and accuracy per epochOpen full-size image ↗

Confusion matrix of the seven emotion classes on the test set — Confusion matrix on the held-out test setOpen full-size image ↗

Try it

Run it locally

No hosted demo yet. The source and README are on GitHub, and these commands start it locally from the repo root.

GitHub repository

terminal

$ pip install "opencv-python>=4.8.0,<4.10" --no-deps
$ pip install -r requirements.txt
$ python -m uvicorn app:app --port 8002 --reload