README.md

metadata

license: mit
task_categories:
  - image-segmentation
tags:
  - wood
  - defect-detection
  - quality-control
  - manufacturing
  - computer-vision
  - semantic-segmentation
size_categories:
  - 1K<n<10K

Dataset Card for Oak Defect Detection

Dataset Description

Dataset Summary

The Oak Defect Detection dataset contains high-resolution line-scan imagery of green rough oak planks with pixel-level annotations for defect detection. The dataset is designed for semantic segmentation tasks to identify and classify defects in wood planks, including BlackRot, Knots, and Stains.

This dataset was collected using line-scan imaging systems and includes both color images and corresponding binary masks for each defect class. The data has been processed and validated to ensure image-mask alignment and quality.

Supported Tasks and Leaderboards

• Semantic Segmentation: The primary task is pixel-level classification of defects in oak planks
• Quality Control: Can be used for automated quality assessment in wood processing
• Defect Classification: Multi-class segmentation with classes: Background, BlackRot, Knot, Stain

Languages

N/A (Image dataset)

Dataset Structure

Data Instances

Each instance consists of:

• Color Image: High-resolution TIFF image (*_Col.tif) captured from line-scan system
• Binary Masks: Per-class binary masks (*_Col_Bin_<ClassName>.tif) for each defect type
• Metadata: Timestamp and position information encoded in filenames

Data Fields

• Images: RGB TIFF format, variable resolution (typical line-scan dimensions)
• Masks: Binary TIFF format, same resolution as corresponding images
• Class Labels:
  • Background (class 0)
  • BlackRot (class 1)
  • Knot (class 2)
  • Stain (class 3)

Data Splits

The dataset includes train/validation/test splits stored in split.json. The splits are designed to ensure:

• Representative distribution of defect classes
• No data leakage between splits
• Sufficient samples for training and evaluation

Dataset Creation

Curation Rationale

This dataset was created to support automated defect detection in wood processing pipelines. The line-scan imagery captures high-resolution details necessary for identifying subtle defects that affect wood quality and value.

Source Data

Initial Data Collection and Normalization

• Data Collection: Images captured using line-scan imaging systems
• Annotation: Binary masks created for each defect class
• Normalization: Defect class names were normalized (e.g., "Black_Rot" → "BlackRot") to ensure consistency
• Filtering: Rare classes below a threshold were filtered to reduce class imbalance

Who are the source language producers?

Industrial line-scan imaging systems in wood processing facilities.

Annotations

Annotation process

Binary masks were created for each defect class, with each mask file corresponding to a specific defect type in a specific image.

Who are the annotators?

Annotations were created as part of the wood processing quality control workflow.

Personal and Sensitive Information

No personal or sensitive information is present in this dataset.

Considerations for Using the Data

Social Impact of Dataset

This dataset supports automated quality control in wood processing, which can:

• Improve efficiency in manufacturing
• Reduce waste through early defect detection
• Standardize quality assessment processes

Discussion of Biases

• Class Imbalance: Some defect classes (e.g., Knot) may be more common than others (e.g., Stain)
• Image Conditions: All images captured under controlled line-scan conditions
• Geographic Bias: Dataset may reflect specific wood sources or processing conditions

Other Known Limitations

• Limited to green rough oak planks
• Line-scan specific imaging conditions
• Some rare defect classes may have been filtered out during preprocessing

Additional Information

Dataset Curators

Created by nrodgers98 for oak defect detection research and applications.

Licensing Information

MIT License