Blasting ore size detection based on efficient dehazing network and multi-dimensional feature fusion.

Area of Science:

• Mining Engineering
• Computer Vision
• Image Processing

Background:

• Ore particle size distribution is crucial for blasting evaluation and crushing energy efficiency.
• Challenges in ore size detection include dense accumulation, non-uniform sizes, dust, and motion-induced target loss.
• Existing computer vision methods require enhancements for robust ore size analysis.

Purpose of the Study:

• To develop an advanced computer vision method for accurate blasting ore size detection.
• To improve the robustness of ore detection models against environmental interferences like dust and wet conditions.
• To enhance the precision and recall rates for detecting various ore sizes, especially fine particles.

Main Methods:

• An efficient dehazing backbone network combining feature attention and a composite scalable backbone was developed.
• A novel feature fusion network integrating convolution and Vmamba sequence models with multi-scale feature fusion was introduced.
• The Dynamic Head was utilized to optimize the target detection head for improved feature fusion and ore discrimination.

Main Results:

• The proposed method, an enhancement of YOLOv8, demonstrated a 7% increase in average precision for eight ore size categories compared to YOLOv8n.
• Mean average precision (mAP50) improved by 7.6% on datasets with dust, smoke, and wet conditions.
• Detection precision for fine ores (<72 mm) increased by 18.8%, with a recall rate rise of 13.8%.

Conclusions:

• The developed method significantly improves blasting ore size detection accuracy and robustness.
• The integration of dehazing, multi-dimensional feature fusion, and optimized detection heads addresses key challenges in ore analysis.
• This approach offers a promising solution for precise ore particle size statistics in industrial applications.