Damaged Board Parts Segmentation Dataset

Project Overview:

Objective

To cultivate a detailed dataset focused on the segmentation of damaged components within various types of boards (e.g., circuit boards, wooden boards, etc.). The primary goal is to facilitate advancements in automation for board inspection systems, damage assessment, and repair recommendations.

Scope

Gather images showcasing a range of board types under different lighting conditions and from varied angles. Every image will contain segmentation masks for both intact and damaged components.

Sources

Engaged in collaborative initiatives with electronics manufacturing units for access to images of defective boards, resulting in a carefully collected and successfully curated assortment of visuals.
Established partnerships with carpentry workshops and lumber mills for samples of damaged wooden boards, contributing to a thoughtfully collected and well-curated dataset.
Crowdsourced submissions from DIY enthusiasts and hobbyists have been meticulously collected and successfully curated for a diverse compilation.
Conducted controlled photography sessions for capturing specific types of damage and wear, resulting in a successfully collected and professionally curated set of images.

Data Collection Metrics

Total Images: 50,000
Electronics Circuit Boards: 25,000
Wooden Boards: 20,000
Other Board Materials (Metal, Plastic, etc.): 5,000

Annotation Process

Stages

Image Pre-processing: Equalize images for resolution and adjust any color imbalances.
Segmentation Annotation: Skilled annotators segment each component, differentiating between damaged and undamaged areas.
Validation: Segmented images are reviewed by industry experts for accuracy.

Annotation Metrics

Total Segmentation Masks: 175,000 (Considering multiple components per image)
Average Annotation Time per Image: 15 minutes (Considering the intricate nature of board components)

Quality Assurance

Automated Verification: Implement early-stage segmentation models to juxtapose their outputs against human annotations, identifying possible misalignments.
Dual Review System: Each segmented image is re-assessed by a second expert to verify consistency and accuracy.
Inter-annotator Agreement: Ambiguous or complex images are reviewed by several annotators, ensuring a consensus on damage delineation.

QA Metrics:

Segmentations Verified using Automated Checks: 25,000 (50% of total images)
Dual-reviewed Segmentations: 15,000 (30% of total images)
Detected and Corrected Inconsistencies: 750 (1.5% of total images)

Conclusion

The Damaged Board Parts Segmentation Dataset emerges as an instrumental tool for industries aiming to harness AI in quality control and board assessment processes. By offering precise segmentation of damaged components, it paves the way for enhanced board inspection systems and more accurate damage evaluations.