Human Skin Detection using Bayesian Classification

Overview

This project implements pixel-wise human skin detection using 2D color histograms in Lab space and Bayesian classification. Trained on 30 skin images (base/skin/) + 30 non-skin images (base/non-skin/), it achieves 85.6% accuracy on test images (base/test/) (threshold 0.4, scale 32).

The project was carried out as part of the computer vision course, in the Master's program in Intelligent Systems and Multimedia at the International Francophone Institute / Vietnam National University. The file Repport/goint_p22-tp1.pdf details the implementation, operation, experimentation, and results of the program.

Dataset Structure

human-skin-detection/
├── src/ → Source files (main.cpp, skin_detection.cpp)
├── include/ → Header files (skin_detection.hpp)
├── base/ → Dataset
│ ├── skin/ → 30 images (only skin visible, rest black)
│ ├── non-skin/ → 30 images (skin black, rest visible)
│ └── test/ → Original images
├── build/ → Object files (.o) generated by Makefile
├── result/ → Output images after detection
├── histogramme/ → 2D histogram images
├── Makefile → Build instructions
└── README.md

Tip: Use GIMP to create masks: Layer → Transparency → Add Alpha Channel → paint non-skin black.

Key features:

• Lab color space conversion
• Histogram smoothing & normalization
• Bayes decision rule
• Morphological post-processing
• Performance evaluation (TP/FP/FN)

Quick start

This README provides Ubuntu (Linux) instructions.

Prerequisites

Before you begin, ensure you have:

• The g++ compiler (GNU C++ Compiler)
• The OpenCV 4 library (C++)

1 . Clone and enter repository

The first things you need to do are cloning this repository and setting up the project environment files

git clone https://github.com/Mongetro/human-skin-detection.git

cd human-skin-detection

2. Build & Compilation

make

• The Makefile uses the src/ and include/ structure.

• Object files are created in build/.

• The executable is human_skin_detection.

3. Run the program with parameters + input image

./human_skin_detection <scale> <threshold> <image_name>

Example :

./human_skin_detection 32 0.4 29.jpg

human_skin_detection : name of the executable generated by make from the C++ code

Parameters

• 32 : scale => Histogram resolution / color space value reduction scale (32 recommended)
• 0.4 : threshold => Bayes threshold (0.4 optimal)
• 29.jpg : image => Filename in base/test/

Results

As indicated in the PDF report goint_p22-tp1.pdf, after running the program, the following specific results are obtained:

• 1 . Display in the terminal

=== PERFORMANCE EVALUATION ===
Reference skin pixels : 21289
True positives        : 20616
False positives       : 2797
False negatives       : 673
Performance           : 85.5933 %
=============================

• 2 . Files saved automatically

  • result/result_image_ImageName.jpg: Final image => skin detected (remains black)
  • histogram/histogram_skin.jpg: histogram/2D (a-b) => histogram of skin pixels
  • histogram/histogram_non_skin.jpg: histogram/2D (a-b) => histogram of non skin pixels

• 3 . OpenCV windows displayed (5 in total)

  • Input Image
  • Reference Skin
  • Detected Skin
  • Skin Mask
  • Histogram - peau
  • Histogram - non_peau

Quick check :

ls result/ histogramme/

Cleanup and recompilation

make clean

make