Skin Segmentation Dataset Builder

An automated tool for converting skin classification datasets into background-free skin segmentation datasets, with a strong focus on facial skin segmentation, while still supporting non-face skin images.

This project is designed for rapid prototyping, research, and production-ready dataset bootstrapping.

---

🚀 What Problem Does This Solve?

Most publicly available skin datasets are classification datasets, structured like:

dataset_original/
├── train/
│   ├── Normal Skin/
│   ├── Acne/
│   └── ...
├── valid/
└── test/

These datasets are excellent for classification, but not usable for segmentation tasks without manual annotation.

Segmentation models require:

• Pixel-level masks
• A different dataset structure
• Careful handling of background pixels

The Core Challenge

In skin segmentation:

• Background pixels dominate images

• Background becomes an overpowering class

• This introduces noise, imbalance, and poor generalization

• Even cropped face images still contain:

  • Hair
  • Eyes
  • Nostrils
  • Clothing
  • Background artifacts

This Tool Solves That By:

✅ Automatically detecting skin regions only ✅ Producing standalone skin images ✅ Producing standalone skin masks ✅ Removing background entirely (no background class) ✅ Preserving train / valid / test splits ✅ Working with:

• Face images (selfies)
• Partial skin images (arms, cheeks, neck, forehead)
• Skin-only datasets

---

🧠 Key Idea

Instead of labeling background pixels as a class, this tool:

Removes all non-skin pixels altogether

This results in:

• Cleaner segmentation datasets
• No background domination
• Better class balance
• Faster convergence during training

---

🧩 How It Works (High Level)

For each image:

1. Attempt face detection (optional)

2. If a face is detected:

   • Prefer face crop (less noise)

3. If no face is detected:

   • Process the full image

4. Detect skin pixels only

5. Generate:

   • Skin-only image
   • Binary skin mask

6. Save outputs in segmentation dataset format

7. Generate colorized mask previews for human inspection (optional)

Face detection is optional and non-blocking.

---

🛠 Tools & Technologies Used

• Python
• OpenCV – image processing
• MediaPipe – optional face detection
• HSV + YCrCb color space filtering – skin detection
• Morphological operations – mask cleanup

No pretrained segmentation model is required.

---

📂 Input Dataset Format (Required)

The tool expects a classification dataset structured as follows:

dataset_original/
├── train/
│   ├── Normal Skin/
│   ├── Acne/
│   └── ...
├── valid/
│   ├── Normal Skin/
│   ├── Acne/
│   └── ...
└── test/
    ├── Normal Skin/
    ├── Acne/
    └── ...

---

📦 Output Dataset Format (Generated)

dataset/
├── images/
│   ├── train/
│   ├── valid/
│   └── test/
├── masks/
│   ├── train/
│   ├── valid/
│   └── test/
├── masks_preview/      # colorized human-readable masks
│   ├── train/
│   ├── valid/
│   └── test/
└── classes.txt

• masks_preview contains colorized masks for visualization only; integer masks in masks are used for training.

🖼 Example Output (Single Sample)

Below is a visual example showing how a single image is transformed by the pipeline.

Original Image	Skin-only Image	Training Mask	Preview Mask

Why the Training Mask Looks Almost Black

The training mask is stored as a single-channel image where each pixel value represents a class ID, not a color.

For example:

• 0 → background (ignored during training)
• 1 → dry skin
• 2 → normal skin
• 3 → oily skin
• ...

Since these values are small integers, they appear very dark or almost invisible when viewed as a normal image.

This is intentional.

Segmentation models require integer-valued masks, not RGB images.
The colorized preview mask is generated purely for human inspection and is never used during training.

Preview Mask Legend

The preview mask uses a fixed color mapping to visualize different skin classes:

• Each color represents a unique skin type or condition
• Colors are assigned automatically and consistently
• Preview masks exist only for debugging and quality checks

---

🔧 Installation

pip install -r requirements.txt

---

▶️ Usage

python build_dataset.py

• The tool will automatically process train, valid, and test splits

• Colorized mask previews will be generated in dataset/masks_preview

• Use CrossEntropyLoss(ignore_index=0) in PyTorch or equivalent in other frameworks

  • Ensures background pixels are ignored during training
  • Only actual skin pixels contribute to learning

---

📜 License

MIT License

---

📬 Contact & Support

Michael Panashe Mudimbu 📧 Email: michaelmudimbu@gmail.com