A synthetic segmentation dataset generator using a 3D modeling framework and raycaster: a mining industry application

Area of Science:

• Computer Vision
• Machine Learning
• Geological Engineering

Background:

• Deep learning, particularly image segmentation, is crucial for industrial efficiency and cost reduction.
• Manual dataset creation for training segmentation models is time-consuming and expensive, especially for specialized applications like deep-level mining.
• Synthetic datasets have emerged as a viable alternative for training accurate segmentation models.

Purpose of the Study:

• To propose and validate a novel synthetic segmentation dataset generator.
• To address the challenges of manual data acquisition in deep-level mining.
• To demonstrate the generator's customizability for diverse applications.

Main Methods:

• Development of a synthetic dataset generator utilizing a 3D modeling framework and raycasting.
• Application of the generator to a deep-level mining case study to create a labeled image dataset.
• Training and validation of image segmentation models using the generated synthetic dataset.

Main Results:

• The generator successfully produced a labeled dataset for deep-level mining environments, eliminating manual data creation.
• Segmentation models trained on the synthetic dataset achieved high accuracy, comparable to models trained on real-world datasets.
• The generator proved to be adaptable for creating datasets for various other applications.

Conclusions:

• The proposed synthetic dataset generator effectively overcomes the limitations of manual data creation for image segmentation.
• This approach significantly reduces the time and cost associated with dataset development in specialized fields like deep-level mining.
• The generator offers a flexible and accurate solution for training deep learning segmentation models across multiple industries.