An Improved Boundary-Aware U-Net for Ore Image Semantic Segmentation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Mineral Processing

Background:

• Accurate ore particle size statistics are crucial for mine operations, relying heavily on precise ore image segmentation.
• Challenges in ore image segmentation include variations in ore size/shape, dust, lighting, complex textures, shadows, and ore adhesion, leading to under-segmentation issues.
• Creating large, labeled datasets for complex ore images is difficult.

Purpose of the Study:

• To propose a novel, multi-task learning network based on U-Net for improved ore image segmentation.
• To address limitations of small datasets and enhance feature extraction capabilities.
• To improve the accuracy of particle size statistics in mining operations.

Main Methods:

• Developed a multi-task learning network utilizing a U-Net architecture.
• Incorporated an improved encoder based on Resnet18 for enhanced feature extraction.
• Designed a decoder with a boundary subnetwork and a mask subnetwork, fused via a boundary mask fusion block (BMFB).

Main Results:

• Achieved 92.07% pixel accuracy, 86.95% Intersection over Union for ore masks (IOU_M), and 52.32% IOU for ore boundaries (IOU_B).
• Reduced the average statistical ore particle size error rate (ASE) to 20.38%.
• Demonstrated improvements over the benchmark U-Net in pixel accuracy (0.65%), IOU_M (1.01%), and IOU_B (5.78%), with a significant reduction in ASE (12.11%).

Conclusions:

• The proposed multi-task learning network effectively addresses challenges in ore image segmentation, particularly under-segmentation.
• The Resnet18-based encoder and dual-subnetwork decoder significantly improve feature extraction and boundary detection.
• The model enhances the reliability of particle size statistics, offering practical benefits for mining operations.