Performance Assessment of Object Detection Models Trained with Synthetic Data: A Case Study on Electrical Equipment...

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Imaging

Background:

• Object detection in industrial settings often faces challenges due to limited real-world image data.
• Manufacturer datasheets offer precise dimensional information valuable for synthetic data generation.

Purpose of the Study:

• To develop and evaluate a data augmentation technique using synthetic images for rigid industrial objects.
• To assess the performance of machine learning detectors trained solely on synthetic data.

Main Methods:

• Generated synthetic images of electrical equipment (lightning rods, potential transformers) using dimensional data from datasheets.
• Trained Viola-Jones (VJ) and You Only Look Once (YOLO) detectors exclusively on synthetic datasets.
• Evaluated detector performance on real-world visible and infrared spectrum images.

Main Results:

• The Viola-Jones detector achieved F1 scores between 38% and 61% on real images.
• The You Only Look Once detector achieved F1 scores below 26% in both spectra.
• The VJ detector's performance suggests the viability of the synthetic data augmentation approach.

Conclusions:

• Synthetic data generation from equipment dimensions is a feasible method for augmenting training datasets.
• The Viola-Jones detector demonstrated superior performance compared to YOLO when trained on synthetic data for this specific task.
• The proposed data augmentation approach shows promise for improving object detection in industrial applications with limited real data.