Card Image Classification Using EfficientNetB3

Project Overview

This project focuses on classifying standard playing cards using a convolutional neural network with EfficientNetB3 as the base model. The dataset includes images of cards from 53 distinct classes, encompassing all suits and ranks, including jokers. The goal is to build a highly accurate classifier that can identify any playing card image.

---

Technical Highlights

• Dataset:

  • 8,154 images representing 53 card types (52 cards + joker)
  • Source: Kaggle - Cards Image Dataset

• Data Pipeline:

  • Images are pre-organized into train, valid, and test directories
  • Image paths and labels are mapped into Pandas DataFrames
  • Image augmentation using Keras' ImageDataGenerator for robust training

• Model Architecture:

  • Base model: EfficientNetB3 pretrained on ImageNet
  • Additional layers: Flatten → Dense(256) → Dropout → Dense(64) → Dropout → Softmax(53)
  • Freezing EfficientNetB3 layers for efficient transfer learning

• Training Configuration:

  • Optimizer: Adam
  • Loss: Categorical Cross-Entropy
  • Epochs: 75
  • Batch size: 40
  • Training and validation accuracy/loss visualized with plots

---

Purpose and Applications

• Card recognition systems for digital casinos and gaming platforms
• Automating card detection in video feeds or camera input
• Training base models for computer vision in object recognition tasks
• Example of applying EfficientNetB3 transfer learning for small custom datasets

---

Installation

Clone the repository:

git clone https://github.com/BhaveshBhakta/Card-Image-Classification-Using-EfficientNetB3.git
cd Card-Image-Classification-Using-EfficientNetB

---

Collaboration

We welcome contributions to make this project more powerful and scalable:

• Add real-time webcam-based prediction
• Deploy via Flask or Streamlit app
• Integrate custom CNN architectures or test other EfficientNet variants
• Explore OCR post-processing for enhanced card detection in gameplay videos