Skin Tone Classification using EfficientNetB0

🎥 Demo Video

This repository contains the Jupyter Notebook and documentation for a deep learning model that classifies images into one of three skin-tone categories: dark, fair, and light.

This model is part of a larger fashion personalization system, as described in the accompanying Phase 1 Report and the academic paper "Optimizing Skin Tone Classification Using EfficientNet for AI-Powered Fashion Personalization."

---

Key Features

• Model: Built using the EfficientNetB0 architecture with transfer learning from ImageNet.
• Dataset: Trained on the SkinTone Dataset from Kaggle, containing 2,100 images
  (1,470 training, 420 validation, 210 test).
• Performance: Final model efficientnet_finetuned_v5.keras achieves 79.05% test accuracy across 3 classes.

---

Tech Stack

• Python
• TensorFlow / Keras
• KaggleHub (for dataset + model loading)
• EfficientNetB0
• scikit-learn (for metrics)

---

Project Resources (Kaggle Hub)

The pre-trained model and dataset are publicly available:

• Pre-trained Model:
  https://www.kaggle.com/models/adityakammati/skintone-images-model

• Dataset:
  https://www.kaggle.com/datasets/adityakammati/skintone-dataset

---

Project Structure

• skin_tone_classification_Traning.ipynb
  Full notebook containing:

  • Data loading
  • Preprocessing
  • EfficientNetB0 fine-tuning
  • Training
  • Saving the model
  • “Run from Here” testing section

• skin_tone_classification_testing_notebook.ipynb
  Notebook for running predictions on new images.

• Skin-Tone-ResearchPaper.pdf
  Complete academic paper.

• HOW_TO_RUN.md
  Step-by-step instructions.

---

How to Use

To run the model, follow the instructions in HOW_TO_RUN.md.