VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect Detection

📄 Related Publication

This repository contains the Official dataset and YOLO-based implementation models for the paper:

“VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect Detection with YOLO Models and Benchmarking”
Published by IEEE

🔗 Paper link: https://ieeexplore.ieee.org/abstract/document/11367538/

---

🧱 Repository Overview

In the sanitary ware ceramics industry, visual quality inspection is essential to ensure product quality and reduce manufacturing defects. Manual inspection is:

• Labor-intensive
• Time-consuming
• Prone to human error

Additionally, ceramic products present unique challenges such as:

• Complex surface geometries
• Highly reflective glaze surfaces

To address these challenges, we introduce VisuCeram, a publicly available dataset specifically designed for ceramic defect detection using YOLO-based object detection models.

---

📊 Dataset Information

The VisuCeram dataset contains:

• 3,265 high-resolution images
• 7 defect categories, including:
  • Surface cracks
  • Glaze imperfections
  • Other ceramic surface defects

📦 Available Dataset Versions

Two versions of the dataset are provided:

1️⃣ VisuCeram-1000x1000.zip

• Image resolution: 1000 × 1000 pixels
• High-quality images
• Intended for visualization and detailed inspection

2️⃣ VisuCeram-500x500-labelled.zip

• Image resolution: 500 × 500 pixels
• Used for all experiments in the paper
• Fully annotated with:
  • Precise bounding boxes
  • YOLO text-format labels
• Annotations created using Label Studio
• Includes detailed class labels for each defect type

This second version is optimized for direct integration with YOLO-based frameworks.

---

🤖 YOLO Models

The repository includes the following YOLO-based architectures:

• YOLOv3-light
• YOLOv4-light
• YOLOv7-light
• VisuCeramNet (proposed model)

All models are located inside the YOLO models folder.

📈 Benchmark Results (mAP)

Model	mAP (%)
YOLOv4-light	87.30%
YOLOv7-light	46.56%
VisuCeramNet	46.10%
YOLOv3-light	41.74%

YOLOv4-light achieved the best performance on the VisuCeram dataset.

---

⚙️ How to Run the Models

To train and test the models, you must use the Darknet framework:

🔗 Darknet repository: https://github.com/AlexeyAB/darknet

Steps:

1. Clone and build the Darknet framework.
2. Place the dataset and model configuration files appropriately.
3. Use the provided YOLO configuration files inside the YOLO models folder.
4. Train or test using Darknet commands.

Please refer to the Darknet documentation for detailed installation and usage instructions.

---

🎯 Purpose of This Repository

• Provide a benchmark dataset for ceramic defect detection
• Support research in industrial visual inspection
• Enable comparison of lightweight YOLO-based detection models
• Promote automation in the sanitary ware ceramics industry

---

📌 Citation

If you use this dataset or code in your research, please cite:

@inproceedings{azad2025visuceram,
  title={VisuCeram: A Comprehensive Dataset for Sanitary Ware Ceramic Defect Detection with YOLO Models and Benchmarking},
  author={Azad, Md Ali and Nahid, Md Mahadi Hasan},
  booktitle={2025 IEEE 7th International Conference on Sustainable Technologies For Industry 5.0 (STI)},
  pages={1--6},
  year={2025},
  organization={IEEE}
}

---