The Pallu Body Recognition Paper: Automatic Classification and Color Changing of Saree Components Using Deep Learning Techniques

 Full FAQ here

🔍 Objective

This research addresses the lack of automation in the saree industry by developing a deep learning-based tool that can:

1. Automatically segment saree components (body and border).

2. Independently recolor each component in a user-friendly and efficient way.

---

🧠 Methodology

The proposed system comprises three main stages:

1. Human Detection & Background Removal

• Faster R-CNN is used to detect whether a human is present in the saree image.

• MODNet (a trimap-free matting method) is used for background removal, especially in human-wearing-saree images.

• MODNet outperforms traditional methods like GrabCut in speed and accuracy.

2. Segmentation of Saree Components

• Two custom-trained Mask R-CNN models are used:

  • One for images with humans (segmenting body, border, hands, face).

  • Another for sarees without humans (segmenting body and border only).

• Data annotated using MakeSense AI, with training on 500 images.

3. Color Changing

• Uses HSV color space instead of RGB, enabling:

  • Independent manipulation of Hue (color) while preserving contrast and brightness (Saturation and Value).

• Dominant hue in a segmented region is detected via histogram peaks and modified as desired.

---

📊 Results

Tested on 50 annotated saree images. Metrics were compared with DeepLabV3+ and BiSeNetV2:

Model	Saree Body Accuracy	Saree Border Accuracy
Proposed Method	93.01%	89.23%
DeepLabV3+	87.45%	86.93%
BiSeNetV2	83.98%	81.21%

• Achieves high precision in pixel-level segmentation.

• Robust even in images with background clutter or overlapping accessories.

• Execution time: ~9 seconds per image.

---

📌 Key Contributions

• First tool for automatic saree component segmentation and independent color customization.

• Enables retailers and designers to visualize color combinations before production.

• Can be generalized to other garments like kurtas, shirts, or kimonos.

---

⚠️ Limitations & Future Scope

• Performance degrades with:

  • Unconventional human poses.

  • Sarees with iridescent textures (e.g., silk).

• Future improvements:

  • Larger and more diverse training datasets.

  • Extension to pattern changes and multi-garment ensembles.