Pokémon Classification Using DenseNet201

Project Overview

This project performs image classification of Pokémon characters using a pretrained DenseNet201 deep learning model. The dataset includes images of 150 different Pokémon classes. The aim is to build a high-accuracy classifier that can automatically recognize and label Pokémon images based on their visual features.

---

Technical Highlights

• Dataset:

  • 6,820 labeled images across 150 Pokémon classes
  • Source: Kaggle - Pokémon Classification Dataset

• Data Pipeline:

  • Custom script to extract image file paths and labels
  • Split into 80% train, 10% validation, 10% test sets
  • Image augmentation via Keras ImageDataGenerator

• Model Architecture:

  • Base model: DenseNet201 pretrained on ImageNet
  • Final layers include GlobalAveragePooling2D and a softmax output layer for 150 classes
  • Freezing first ~675 layers for transfer learning
  • Optimizer: Adam with a learning rate of 0.001

• Training:

  • Run for 30 epochs with categorical cross-entropy loss
  • Evaluation includes training vs validation accuracy and loss visualization

• Performance:

  • Achieved strong training/validation accuracy with DenseNet201
  • Suitable for real-world Pokémon classification or gaming datasets

---

Purpose and Applications

• Automated Pokémon classification for games and AR/VR applications
• Building educational datasets for computer vision and deep learning
• Demonstrating transfer learning using pretrained CNNs
• Framework adaptable to other multi-class image classification tasks

---

Installation

Clone the repository:

git clone https://github.com/BhaveshBhakta/Pokemon-Classification-Using-DenseNet201.git
cd Pokemon-Classification-Using-DenseNet20

---

Collaboration

We welcome contributions to enhance this project. You can:

• Integrate more CNN architectures like EfficientNet or ResNet
• Convert to streamlit/flask web app for live image upload predictions
• Extend the project to classify evolution stages or types (Fire, Water, etc.)