Comparative Analysis of Pretrained ImageNet Models for Static Hand Gesture Recognition

📌 Overview

This repository contains a detailed comparative analysis conducted on 15 state-of-the-art CNN pretrained ImageNet models for character-level static hand gesture recognition using the Sign Language Gesture Images Dataset.

🎯 Objectives

• Evaluate and compare the performance of pretrained ImageNet models.
• Identify the most accurate, efficient, and robust models for static hand gesture recognition.
• Provide insights to guide model selection for real-world Human-Computer Interaction (HCI) applications.

🚀 Models Evaluated

EfficientNet Family	ResNet Family	VGG Family	MobileNet Family	DenseNet Family	ConvNeXt Family
EfficientNetV2L	ResNet50	VGG16	MobileNetV2	DenseNet121	ConvNeXtBase
EfficientNetB0	ResNet101	VGG19	MobileNetV3Large	DenseNet169	ConvNeXtXLarge
EfficientNetV2S	ResNet152			DenseNet201

🗃️ Dataset

• Name: Sign Language Gesture Images Dataset
• Size: 55,500 images (37 classes, including letters, numbers, and space)
• Image Dimension: 50×50 pixels

🛠️ Repository Structure

CAPIMSHGR/
├── code/                   # All model implementations
│   ├── [model]_[author].ipynb
│   └── ...
├── dataset/                # Dataset files
│   └── dataset.zip
├── docs/                   # Reports and documentation
│   └── readme.md
└── README.md               # Project overview (you are here)

🚧 Instructions for Contributors

• Download notebook from Kaggle.
• Rename the notebook to [model]_[author].ipynb format (e.g., vgg19_adib.ipynb).
• Place notebooks into the code folder.
• Commit each notebook separately for clarity and maintainability.

⚠️ Please do not modify or delete other contributors' files.

🔬 Methodology

• Preprocessing: Data normalization, augmentation, stratified data splitting (60%-20%-20%).
• Training Strategy: Transfer learning, fine-tuning final layers, and extensive model evaluation.
• Evaluation Metrics: Accuracy, Precision, Recall, F1-score.
• Interpretability Techniques: Grad-CAM, LIME, t-SNE visualization.

📈 Key Findings

• Highest Accuracy: ConvNeXtXLarge (99.63% Top-1 accuracy).
• Balanced Choice: EfficientNetB0, MobileNetV2 (Optimal for resource-constrained environments).

Recommended for	Model(s)
Highest Accuracy	ConvNeXtXLarge, VGG19
Best Efficiency (Edge Devices)	EfficientNetB0, MobileNetV2
Balanced Performance	ResNet50, EfficientNetV2S

🎨 Visual Insights

• Grad-CAM heatmaps: Highlight crucial areas for gesture recognition.
• t-SNE plots: Demonstrate clear clustering between gesture classes.

📖 Full Report

• The detailed PDF report is available in the docs directory.

💬 Contributors

• Adib Sakhawat - 210042106
• Md Hasibur Rahman - 210042107
• Minhajul Abedin - 210042148
• Nabila Islam - 210042111
• Nazifa Tasneem - 210042114

🙌 Acknowledgments

Special thanks to the Kaggle platform for providing computational resources for the experiments.

---

✨ Happy Exploring!

For any queries or further collaboration, please contact the contributors directly.