Related Experiment Videos
An Image Dataset Quality Evaluation System for Industrial Object Detection Tasks
Shengguo Zhu1, Yunxi Sun1, Enhui Lu1
1School of Mechanical Engineering, Yangzhou University, Yangzhou 225127, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a framework to evaluate industrial object detection dataset quality. Higher dataset quality scores correlate with improved model accuracy, offering practical guidelines for better industrial imaging.
Area of Science:
- Computer Vision
- Machine Learning
- Industrial Automation
Background:
- Industrial object detection model performance heavily relies on training data quality.
- Current focus on imaging schemes and algorithms overlooks systematic dataset evaluation.
Purpose of the Study:
- To propose a comprehensive dataset quality evaluation framework for industrial object detection.
- To provide quantifiable metrics and practical guidelines for enhancing industrial image datasets.
Main Methods:
- Developed a framework with four dimensions (acquisition environment, image quality, dataset scale, annotation quality) and thirteen indices.
- Employed Taguchi orthogonal experiments for dimension weighting and normalization for score bias reduction.
- Validated the framework on multiple datasets using YOLOv12n and RT-DETR-R18 models.
Main Results:
- Observed a significant positive correlation between proposed dataset quality scores and detection accuracy (PLCC/SRCC/Kendall's tau up to 0.922).
- Weight optimization further enhanced the correlation, demonstrating the framework's effectiveness.
- Sensitivity analysis and ablation studies confirmed the framework's robustness and necessity.
Conclusions:
- The proposed framework offers a quantifiable method for assessing industrial image dataset quality.
- The study provides practical acquisition guidelines to improve dataset quality and subsequent model performance.