Pokemon Multi-Layer Perceptron Classifier 🎮🤖

A machine learning project that classifies Pokemon images using a Multi-Layer Perceptron (MLP) neural network with optimized color histogram features.

📋 Project Overview

This project implements a comprehensive machine learning pipeline to classify 10 different Pokemon species using computer vision techniques and neural networks. The system automatically determines the optimal MLP architecture and histogram configuration through systematic experimentation. Key Features

Automated Architecture Selection: Tests 9 different MLP structures to find optimal configuration Feature Optimization: Evaluates multiple color histogram sizes for best performance Comprehensive Evaluation: Implements accuracy, precision, recall, and F1-score metrics Visual Analysis: Generates confusion matrices to identify classification patterns

🎯 Pokemon Classes

The classifier distinguishes between 10 Pokemon:

Bulbasaur Meowth Mew Pidgeot Pikachu Snorlax Squirtle Venusaur Wartortle Zubat

🛠️ Technologies Used

Python 3.x scikit-learn: MLP implementation and evaluation metrics OpenCV: Image processing and feature extraction NumPy: Numerical computations Matplotlib: Visualization tqdm: Progress tracking

📊 Methodology

1. Data Pipeline

Preprocessing: Images resized to 150x150 pixels Split Strategy:

Training: 70% Validation: 10% Testing: 20%

2. Feature Extraction

Utilizes color histograms as feature vectors:

Extracts RGB color distribution Tests histogram bins: 4x4x4, 6x6x6, and 8x8x8 Normalizes features for consistent scaling

3. Architecture Optimization

Evaluates 9 different MLP configurations:

Single hidden layer: 3 variations based on neuron count formulas Two hidden layers: 3 configurations Three hidden layers: 3 deep network structures

Neuron count calculations:

Formula 1: (input_size + output_size) / 2 Formula 2: (2/3 × input_size) + output_size Formula 3: 80% of input_size

4. Model Training

Activation Function: ReLU Optimizer: Adam Early Stopping: Enabled to prevent overfitting Maximum Iterations: 1500

📈 Results

The system automatically identifies:

Optimal MLP architecture for the dataset Best performing histogram configuration Most frequently confused Pokemon pairs Detailed performance metrics for each configuration

Performance Metrics

Accuracy: Overall classification correctness Precision: Positive prediction accuracy Recall: True positive identification rate F1-Score: Harmonic mean of precision and recall

🚀 Getting Started

Prerequisites bashpip install numpy opencv-python scikit-learn matplotlib tqdm Dataset Structure Organize your Pokemon images as follows:

Dataset/

└── PokemonData/ ├── Bulbasaur/ │ ├── image1.jpg │ └── ... ├── Meowth/ │ └── ... └── [Other Pokemon folders] Running the Classifier bashpython Multi-Layer-Perceptron-Classifier.py Expected Output

Data split statistics

Neuron count calculations for each formula Performance metrics for each MLP structure Optimal structure identification Results for different histogram sizes Final confusion matrix visualization Most confused Pokemon pair identification

📊 Sample Output

DETERMINING OPTIMAL MLP STRUCTURE

The Optimal MLP Structure is: (154, 154) With validation accuracy: 0.8567

CLASSIFICATION METRICS FOR MLPCLASSIFIER

The Accuracy Score for this model is: 0.8734 The Precision Score for this model is: 0.8756 The Recall Score for this model is: 0.8734 The f1 Score for this model is: 0.8721

🔍 Key Insights

Feature Engineering Color histograms prove effective for Pokemon classification, capturing distinctive color patterns of each species while remaining computationally efficient. Architecture Selection The systematic approach to architecture selection ensures optimal performance without manual hyperparameter tuning. Confusion Analysis The confusion matrix reveals which Pokemon share similar visual characteristics, providing insights for potential feature improvements.

🎓 Learning Outcomes

This project demonstrates:

1. Machine Learning Pipeline: End-to-end ML workflow implementation
2. Neural Network Design: Understanding of MLP architecture decisions
3. Computer Vision: Practical feature extraction techniques
4. Model Evaluation: Comprehensive performance assessment
5. Code Organization: Clean, documented, and modular programming

🔄 Future Improvements

Implement CNN for improved accuracy Add data augmentation techniques Expand to more Pokemon species Create web interface for real-time classification Implement transfer learning approach Add cross-validation for more robust evaluation

📄 License

This project is for educational purposes. Pokemon characters are property of Nintendo/Game Freak.

👤 Author

Ben Fricker

Computer Science Student at University of Wollongong Majoring in AI & Big Data and Cybersecurity LinkedIn GitHub

🤝 Acknowledgments

University of Wollongong for project guidance scikit-learn documentation for MLP implementation details Pokemon dataset contributors

This project is part of my journey in exploring machine learning and computer vision applications.