Yoga Pose Classification Using Transfer Learning

Overview

Yoga is a practice that originated in ancient India, aimed at balancing the mind and body through meditation, exercise, and regulated breathing. Hatha Yoga, a type of physical yoga, consists of postures (asanas) performed in a continuous sequence. Correct identification of yoga poses is essential to ensure proper practice and alignment with individual needs.

This repository provides an implementation of yoga pose classification from images using a transfer learning approach. The model classifies yoga poses from images of five different asanas, leveraging pre-trained deep learning models to achieve high accuracy.

Research paper link : https://ijirt.org/publishedpaper/IJIRT167821_PAPER.pdf

Summary

In this project, a total of 1551 images representing 5 distinct yoga postures were used. The images were resized to optimize computation. Transfer learning was employed, utilizing 10 different pre-trained models for classification.

Dataset

• Total Images: 1551
• Yoga Poses: 5 distinct yoga asanas
• Image Preprocessing: All images were resized for consistency and computational efficiency.

Models

The following pre-trained models from popular deep learning architectures were used for classification:

1. VGG16
2. VGG19
3. InceptionV3
4. DenseNet201
5. ResNet50V2
6. ResNet101V2
7. ResNet152V2
8. MobileNet
9. MobileNetV2
10. InceptionResNetV2

Best Model

• VGG16 performed the best with a validation accuracy of 94.47%.

Requirements

To replicate the results, you will need the following libraries:

• TensorFlow
• Keras
• NumPy
• seaborn
• Matplotlib (for visualizations)
• Scikit-learn (for model evaluation metrics)

You can install the dependencies by running:

pip install -r requirements.txt

Implementation

follow these steps:

1. Clone the repository:

   git clone https://github.com/sumony2j/Transfer-Learning.git

2. Navigate to the project directory:

   cd Transfer-Learning

3. The Dataset is in /Data directory in zip format (Yoga_New_Dataset_Resized.zip)
4. Navigate to the Code implementation directory:

   cd Implementation

Results

• VGG16 achieved the highest accuracy of 94.47% on the validation set.
• Model performance metrics (accuracy, precision, recall, F1-score) will be printed for each model.

Research Paper

This repository contains the implementation of my published research paper on Yoga Pose Classification Using Transfer Learning. The paper presents a comparative analysis of 10 different deep learning models for classifying yoga poses from images.

For more details, you can refer to the published paper:

• Title: Yoga Pose Classification Using Transfer Learning
• Author: Sumon Singh,Deepali Handgar
• Published In: International Journal of Innovative Research in Technology (https://ijirt.org/Article?manuscript=167821)
• DOI/Link: https://ijirt.org/publishedpaper/IJIRT167821_PAPER.pdf

Future Work

• Fine-tuning: Fine-tune pre-trained models on the yoga dataset for improved performance.
• Additional Poses: Expand the dataset to include more yoga poses.

Acknowledgments

We would like to thank the open-source community for providing pre-trained models and tools that made this project possible.