Pokémon Image Classifier
A convolutional neural network trained to identify four Pokémon species from images achieving 88–96% accuracy using TensorFlow, Keras, and a custom preprocessed dataset.
Introduction
For our final Machine Learning course assignment, the team and I set out to combine the world of Pokemon with the power of Convolutional Neural Networks. The goal: build an image recognition system capable of distinguishing between four unique Pokemon species with high accuracy.
CNNs are well-suited for image classification tasks due to their ability to learn spatial hierarchies of features, making them a natural fit for differentiating visually distinct (and sometimes surprisingly similar) Pokémon.
The four Pokémon species selected for classification
Tools & Technologies
TensorFlow / Keras
CNN model implementation
Scikit-learn
Train/test split & metrics
OpenCV & NumPy
Image preprocessing
Matplotlib
Results visualization
Google Colab
Development environment
Kaggle
Image dataset hosting
Setup & Methodology
For each Pokémon we collected between 270 and 370 images sourced online, uploaded to Kaggle, and imported into Google Colab. The dataset was split 80/20 into training and testing samples, with pixel values normalized by dividing by 255.
The model architecture was the most time-intensive part, requiring multiple iterations to find the right number of layers for optimal accuracy without overfitting. We settled on a 7-layer CNN, trained for up to 15 epochs with early stopping when accuracy began to decline.
Model Architecture (7 Layers)
Conv2D ×2
Detect subtle visual features with increasing depth
MaxPooling2D ×2
Extract the most prominent features from pixel arrays
Dense ×2
Classify images using convolutional output
Dropout
Prevent overfitting during training
Results
~96%
Peak test accuracy
88–96%
Accuracy across runs
4
Pokémon species classified
The most common misclassification was Charmander being identified as Pikachu, likely due to shared visual traits like a tail and similar warm skin tones. The confusion matrix and epoch graph below illustrate the model's performance in detail.
Confusion matrix
Accuracy across epochs
Conclusion
This project demonstrates that a CNN can reliably classify Pokémon images with high accuracy as long as the species don't share too many visual similarities. The 88–96% accuracy range held consistent across multiple runs with different dataset samples.
Future improvements would include expanding to more Pokémon species, increasing image count and quality per class, and experimenting with transfer learning using pretrained models like ResNet or MobileNet for even higher accuracy.