Undergraduate Research

CNN School Face Recognition System

A team-built computer-vision system for recognizing uniformly dressed pupils in school surveillance footage.

Three university researchers reviewing an anonymized computer-vision surveillance system

Computer Vision & School Safety

Project overview

Developed by a three-person undergraduate research team, this project explored how a convolutional neural network could support student safety and incident investigation at KNUST Basic School. The pipeline turns CCTV footage into labeled face data, trains a multiclass CNN, and applies the model to video frames for identity prediction.

Final-year undergraduate research completed by a team of three.

95.45%

Reported validation accuracy

Source: Thesis abstract

~20 GB

CCTV footage processed

Source: Thesis abstract

Identity classes

Source: Model architecture

882,842

Model parameters

Source: Keras model summary

The problem

Uniform school clothing makes visual identification from surveillance footage difficult, limiting how quickly schools can investigate bullying, vandalism, unauthorized access, and other safety incidents.

Approach

1
Process approximately 20 GB of CCTV footage and extract face images for individual pupil classes.
2
Resize faces to 112 × 92 pixels, convert them to grayscale, normalize the values, and augment the training data.
3
Train a four-block convolution and max-pooling network with a 128-unit dense layer, dropout, and a 26-class softmax output.
4
Evaluate predictions with accuracy and loss curves, a confusion matrix, and classification metrics.
5
Deploy the saved Keras model with OpenCV to detect faces, predict identities, and annotate video frame by frame.

Outcomes and insights

The thesis reports 95.45% validation accuracy and successful recognition when pupils were sufficiently close to the camera.
Testing identified camera distance and placement as important practical limits on recognition quality.
The work demonstrates an end-to-end research pipeline from CCTV data preparation through model training and video deployment.
The prototype provides a foundation for safer, carefully governed real-time monitoring in educational environments.

Research evidence

Selected technical figures from the thesis and implementation.

CNN training and validation accuracy and loss curves — Training history used to compare learning and validation behaviour across epochs.

Confusion matrix for the pupil identity classes — Confusion matrix showing class-level predictions during model evaluation.

Keras summary of the convolutional neural network — The implemented CNN contains four convolution blocks and 882,842 trainable parameters.

Diagram explaining convolutional feature learning and classification — Conceptual CNN pipeline from input images through feature learning to classification.

This public case study intentionally excludes identifiable pupil photographs and surveillance frames. Only technical figures and privacy-safe imagery are shown.

Project overview

Final-year undergraduate research completed by a team of three.

95.45%

Reported validation accuracy

Source: Thesis abstract

~20 GB

CCTV footage processed

Source: Thesis abstract

Identity classes

Source: Model architecture

882,842

Model parameters

Source: Keras model summary

Approach

Process approximately 20 GB of CCTV footage and extract face images for individual pupil classes.

Resize faces to 112 × 92 pixels, convert them to grayscale, normalize the values, and augment the training data.

Train a four-block convolution and max-pooling network with a 128-unit dense layer, dropout, and a 26-class softmax output.

Evaluate predictions with accuracy and loss curves, a confusion matrix, and classification metrics.

Deploy the saved Keras model with OpenCV to detect faces, predict identities, and annotate video frame by frame.

Outcomes and insights

The thesis reports 95.45% validation accuracy and successful recognition when pupils were sufficiently close to the camera.

Testing identified camera distance and placement as important practical limits on recognition quality.

The work demonstrates an end-to-end research pipeline from CCTV data preparation through model training and video deployment.

The prototype provides a foundation for safer, carefully governed real-time monitoring in educational environments.

Research evidence

Selected technical figures from the thesis and implementation.

This public case study intentionally excludes identifiable pupil photographs and surveillance frames. Only technical figures and privacy-safe imagery are shown.