A critical challenge in training neural networks for mining is the lack of diverse data. In the primary study, an initial set of 1,980 original images was collected. To improve generalization and prevent overfitting, various were applied: Geometric Transformations : Image rotation (randomly ±90plus or minus 90 Photometric Adjustments : Random luminance changes (up to ) to simulate varying lighting underground.
) and real-time processing speeds, outperforming traditional YOLO architectures in underground mining environments. 1. Introduction 11265.rar
The use of the expanded 11,265-sample dataset was foundational to achieving a model that is both accurate and fast enough for industrial application. Through transfer learning, the algorithm has been successfully applied to underground image segmentation, verifying its reliability as an automated solution for the coal industry. A critical challenge in training neural networks for
FPS increase, enabling real-time deployment on conveyor belt systems. 5. Conclusion Through transfer learning