Garbage Classification with Feature Selection via Genetic Algorithm

Project Overview

This project implements a robust pipeline for image-based garbage classification. It leverages transfer learning for feature extraction, followed by an advanced feature selection technique using a Genetic Algorithm (GA) to optimize a Support Vector Machine (SVM) classifier. The goal is to enhance model performance and generalization by identifying the most discriminative subset of features from pre-trained convolutional neural networks.

Methodology

The implemented solution follows a multi-stage, hybrid approach:

1. Data Preparation & Augmentation

• The dataset is partitioned into dedicated training and validation sets.
• To combat overfitting and improve model generalization, we employ data augmentation using Keras' ImageDataGenerator. This introduces variability in the training data through random transformations (e.g., rotations, shifts, flips, zooms).

2. Deep Feature Extraction

• We utilize pre-trained, state-of-the-art convolutional neural networks (CNNs) as feature extractors. Models such as MobileNet and Xception (trained on ImageNet) are evaluated for this purpose.
• The final classification layers are removed, and a Global Average Pooling (GAP) layer is appended to the base model. This step reduces the spatial dimensions of the extracted feature maps, producing a compact, 1D feature vector for each input image.
• These high-level, deep features are extracted for both the training and validation datasets, forming our initial feature space.

3. Evolutionary Feature Selection

To reduce dimensionality and eliminate redundant features, a custom Genetic Algorithm is implemented for feature selection.

• Population Initialization: A population of individuals is generated, where each individual is a binary-encoded vector. Each bit represents the presence (1) or absence (0) of a specific feature in the subset.
• Fitness Evaluation: The fitness of each individual (feature subset) is calculated by training an SVM classifier exclusively on the selected features. The primary fitness metric is the classifier's accuracy on the validation set.
• Parent Selection: A tournament selection strategy is employed. For each new offspring, two random individuals are chosen from the population, and the one with the higher fitness score is selected as a parent.
• Crossover: A single-point crossover operation is performed on pairs of selected parents. A random crossover point is chosen, and genetic material is swapped between parents to produce two new offspring.
• Mutation: Random bit flips are applied to the offspring with a low probability, introducing genetic diversity into the population and preventing premature convergence to a local optimum.
• Iteration: This process of selection, crossover, and mutation repeats for a set number of generations, evolving the population towards an optimal feature subset.

4. Final Classification

The feature subset identified by the genetic algorithm as optimal is used to train a final SVM classifier. This model is then evaluated on the test set to report final performance metrics (e.g., accuracy, precision, recall, F1-score).

Key Features

• Hybrid Architecture: Combines the representational power of deep CNNs with the computational efficiency and strong theoretical foundations of SVMs.
• Optimized Feature Space: The genetic algorithm performs intelligent feature selection, leading to a model that is potentially more accurate, less complex, and more interpretable.
• Robustness: Data augmentation and careful validation ensure the model generalizes well to unseen data.

Key libraries include:

• tensorflow / keras
• scikit-learn
• numpy
• opencv-python
• matplotlib

Usage

1. Prepare Data: Organize your image dataset into train and validation directories, with subdirectories for each class.
2. Extract Features: Run the feature extraction script (feature_extractor.py) using your chosen pre-trained model (e.g., MobileNet). This will generate .npy files containing the features and labels.
3. Run Genetic Algorithm: Execute the genetic algorithm script (genetic_algorithm.py) to find the optimal feature subset. This will output a binary mask of the selected features.
4. Train Final Model: Use the classifier.py script to train and evaluate the final SVM model on the selected features.

Results

The model achieves competitive accuracy on the garbage classification task. The evolutionary feature selection step consistently leads to a significant reduction in the number of features used (often >50%) while maintaining or improving classification performance compared to using the full feature set.

Future Work

• Experiment with other pre-trained models (e.g., EfficientNet, ResNet).
• Incorporate multi-objective optimization in the GA to simultaneously maximize accuracy and minimize feature count.
• Develop an end-to-end deep learning model that integrates the feature selection process directly into the network architecture.