FabricSpotDefect: An annotated dataset for identifying spot defects in different fabric types.

Area of Science:

• Computer Vision
• Machine Learning
• Textile Quality Control

Background:

• Fabric defect detection is crucial for textile quality control.
• Existing datasets may lack diversity in fabric types and defect complexities.
• Manual inspection is labor-intensive and prone to errors.

Purpose of the Study:

• Introduce the FabricSpotDefect dataset, the first specifically for challenging computer vision models in fabric spot detection.
• Provide a diverse and realistic dataset for training and validating machine learning models.
• Facilitate the development of advanced AI tools for automated textile inspection.

Main Methods:

• Collected 1014 raw images with 3288 manual annotations of various fabric spot defects.
• Applied six augmentation techniques to create 2300 enhanced images.
• Annotated data using bounding boxes and polygons for precise defect localization.
• Formatted dataset in YOLOv8 and COCO formats for compatibility with ML frameworks.

Main Results:

• The dataset includes diverse fabrics (cotton, silk, denim) and challenging spot types (stains, rust, discolorations).
• Annotations were performed on original images to ensure real-world relevance.
• Augmented data increases diversity, reduces overfitting, and improves model robustness.
• The dataset is readily available in Roboflow for research and development.

Conclusions:

• The FabricSpotDefect dataset is a valuable, unique resource for advancing computer vision in textile defect detection.
• It supports the development of AI-powered quality control systems for various textile applications.
• The dataset's real-world conditions and diverse examples make it ideal for robust model training.