VTC-Net: A Semantic Segmentation Network for Ore Particles Integrating Transformer and Convolutional Block Attention...

Area of Science:

• Mineral Processing
• Computer Vision
• Image Segmentation

Background:

• Accurate ore particle size distribution is crucial for optimizing mineral processing operations.
• Existing image segmentation models struggle with ore particle clusters due to adhesion, occlusion, and multi-scale variations, leading to inaccurate results.

Purpose of the Study:

• To develop a novel semantic segmentation model, VTC-Net, for high-precision visual-based online particle size analysis in mineral processing.
• To overcome the limitations of current models in handling complex ore particle characteristics.

Main Methods:

• Proposed VTC-Net model utilizing VGG16 backbone and Transformer modules for global context.
• Integrated Convolutional Block Attention Module (CBAM) to focus on critical adhesion edges.
• Employed BatchNorm layers for stable training.

Main Results:

• VTC-Net achieved superior performance over UNet and DeepLabV3 on ore image datasets.
• Key metrics demonstrated VTC-Net's effectiveness: MIoU of 89.90% and pixel accuracy of 96.80%.
• Ablation studies validated the contribution of each VTC-Net module.

Conclusions:

• VTC-Net significantly improves segmentation robustness and precision for ore particle analysis under complex conditions.
• The model accurately identifies ore contours and adhesion areas, enhancing operational optimization.