MATLAB project developed for the Knowledge and Reasoning course at the Polytechnic Institute of Coimbra (ISEC).
The goal is to design, train, and evaluate feedforward neural networks capable of classifying geometric shapes from images.
- Image preprocessing (resizing, grayscale conversion, binarization)
- One-hot encoding for target representation
- Training and testing of feedforward networks with various topologies
- Comparative analysis of:
- Training and activation functions
- Data division ratios (train/validation/test)
- Model generalization and overfitting
- Confusion matrix generation and performance visualization
- Graphical User Interface (GUI) for interactive model training and classification
- MATLAB
- Deep Learning Toolbox
- Image Processing Toolbox
The best models achieved 100% accuracy on training datasets and strong generalization across different image sets.
Further experiments revealed the impact of training parameters and dataset diversity on classification performance.
An interactive MATLAB GUI allows the user to:
- Configure network parameters (neurons, layers, functions)
- Train or import models
- Load or draw new images
- Classify shapes and visualize results
Developed by Ana Rita Pessoa and João Francisco Claro (ISEC, 2024/2025).
This project explores neural networks for image-based classification, bridging concepts of machine learning, pattern recognition, and data analysis.