This project is a final laboratory assignment for the course Audiovisual Perception and Cognition, taught by Prof. Elvira García Guzman, within the Audiovisual Systems Engineering Degree. The project was developed by Xavier Riera, Aina Gutiérrez, Alexandra Anashkina and Pau Noguera. The objective of the project is to explore audiovisual perception and visual feature extraction through the use of Gabor filters and neural networks to classify images into two categories: birds and airplanes. The complete pipeline is implemented in MATLAB.
The system performs binary image classification using images extracted from the CIFAR dataset, focusing on distinguishing birds vs. airplanes. Inspired by early stages of the human visual system, the project applies Gabor filters to extract orientation and frequency-sensitive features before feeding the processed data into a Multilayer Perceptron (MLP) classifier. The workflow consists of:
- Image preprocessing and feature extraction using Gabor filters.
- Dimensionality reduction to optimize computational efficiency.
- Neural network training using backpropagation.
- Validation and performance analysis to evaluate generalization and avoid overfitting.
The Gabor filter plays a key role in detecting relevant visual features such as edges and textures. The main parameters used are:
- Kernel size: Determines the spatial extent of the filter.
- Lambda (λ): Controls the wavelength of the sinusoidal component, affecting spatial frequency sensitivity.
- Phi (φ): Defines the phase offset of the sinusoidal function.
- Sigma (σ): Sets the width of the Gaussian envelope, controlling how localized or diffuse the filter is.
- Gamma (γ): Aspect ratio of the filter, influencing elongation perpendicular to the sinusoidal direction.
Validation is performed on a dataset not seen during training to evaluate the model’s ability to generalize. The results show that the combination of Gabor-based feature extraction and MLP classification is effective in distinguishing between birds and airplanes, while also illustrating common challenges such as overfitting and dataset variability. Visualization tools (error plots, classification histograms, and prediction grids) are used to interpret the behavior of the model and the quality of the learned features.
We acknowledge Prof. Elvira García Guzman for providing the MATLAB template and core functions used in this project as part of the Audiovisual Perception and Cognition course. The template served as a starting point for implementing and extending the image classification pipeline, including feature extraction, neural network training, and performance evaluation.