Skin Color Detection Using Chromaticity

---

Overview

This project implements a statistical skin colour detection system using chromaticity space. The model learns the distribution of skin pixel colours from a labelled dataset and uses Mahalanobis distance to classify each pixel in a new image as either skin or non-skin.

The key idea is that converting raw RGB values to chromaticity coordinates removes the effect of lighting, making skin detection more consistent across different conditions.

---

How It Works

Raw RGB values change with lighting — the same skin under bright or dim light gives very different R, G, B numbers. Chromaticity normalises this by dividing each channel by the total intensity:

x = R / (R + G + B)
y = G / (R + G + B)

Skin pixels form a tight cluster in this 2D chromaticity space regardless of brightness. We model that cluster with a Gaussian distribution (mean + covariance), then classify any new pixel by measuring how far it sits from that cluster.

---

Dataset

Pratheepan Skin Dataset — cs-chan.com

Images collected from Google covering a range of skin tones, lighting conditions, and backgrounds. Each image comes with a ground truth mask where white pixels indicate actual skin regions.

---

Folder Structure

project/
│
├── data/
│   ├── images/           # Full original photos
│   ├── skin_cropped/     # Manually cropped skin regions (used for training)
│   └── ground_truth/     # Binary masks (white = skin, black = non-skin)
│
├── out/                  # All output images saved here
│
├── skin_detection.ipynb  # Main notebook
└── README.md

Make sure the out/ folder exists before running, or create it with os.makedirs('out', exist_ok=True).

---

Notebook Walkthrough

Cell	Description
1	Import libraries — cv2, numpy, matplotlib, os
2	Set dataset folder paths
3	Load and count image files
4	Visualise sample images alongside their cropped skin regions
5	Visualise chromaticity x and y channels on sample images
6	Extract skin pixel chromaticity values from cropped skin images
7	Build the skin model — compute mean and covariance matrix
8	Collect non-skin pixel samples from full images
9	Build the non-skin model for comparison
10	Plot skin vs non-skin pixel distributions in chromaticity space
11	skin_detection() function + threshold comparison across values 2, 5, 10, 15, 20
12	evaluateSkinNonSkinDetection() — full visual output with ground truth comparison

---

Detection Pipeline

Input Image
     │
     ▼
Convert to Chromaticity (x, y)
     │
     ▼
Compute Mahalanobis Distance from Skin Mean
     │
     ▼
Apply Threshold  ──────────────────────────┐
     │                                     │
  distance < threshold             distance >= threshold
     │                                     │
  SKIN (1)                            NON-SKIN (0)

---

Output

For each image the notebook produces a 5-column visual:

Column	Description
Original Image	The unmodified input photo
Binary Mask	White = detected skin, Black = non-skin
Skin Only	Original colours kept only where skin is detected
Skin + Grey BG	Skin in real colour, non-skin converted to greyscale
Ground Truth	The correct answer mask from the dataset

All outputs are saved to the out/ folder as .png files.

---

Key Functions

# Convert RGB image to chromaticity coordinates
convert_to_chromaticity(img)
    → returns x, y  (arrays of same shape as image)

# Detect skin pixels using Mahalanobis distance
skin_detection(image, meanValue, conInverse, Threshold)
    → returns binary mask  (1 = skin, 0 = non-skin)

# Produce skin-only and combined visualisation
evaluateSkinNonSkinDetection(image, mask)
    → returns skinOnly, combinedImg

---

Requirements

pip install opencv-python numpy matplotlib

Library	Version	Purpose
opencv-python	4.x	Image loading, colour conversion
numpy	any	Matrix operations, statistics
matplotlib	any	Plotting and visualisation

---

Running the Notebook

# Clone or download the project folder
# Place the Pratheepan dataset into data/ as shown above
# Launch Jupyter
jupyter notebook skin_detection.ipynb
# Run all cells top to bottom

---

Results Saved

File	Description
out/sample_images_cropped_and_original.png	Dataset sample preview
out/chromaticity_conversion_samples.png	Chromaticity x/y visualisation
out/skin_chromaticity_distribution.png	Skin pixel cluster plot
out/skin_non_skin_chromaticity_distribution.png	Skin vs non-skin scatter plot
out/skin_detection_result.png	Final 5-column detection output

---

References

W.R. Tan, C.S. Chan, Y. Pratheepan and J. Condell — A Fusion Approach for Efficient Human Skin Detection, IEEE Transactions on Industrial Informatics, vol.8(1):138-147, 2012.