YOLOv8-Coal: a coal-rock image recognition method based on improved YOLOv8

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Coal-rock image recognition faces challenges like misdetection and omission due to poor lighting, defocus, and occlusion.
• Existing methods struggle to achieve high accuracy and processing speed in complex industrial environments.

Purpose of the Study:

• To introduce YOLOv8-Coal, an enhanced object detection model for improved coal-rock image recognition.
• To address limitations of existing methods by enhancing feature extraction, feature fusion, and model efficiency.

Main Methods:

• Utilized YOLOv8 as the base architecture.
• Incorporated Deformable Convolution Network v3 for adaptive feature extraction.
• Implemented a Polarized Self-Attention module for refined feature fusion.
• Introduced a C2fGhost module to reduce model complexity and computational load.

Main Results:

• YOLOv8-Coal achieved significant improvements in AP50 (77.7%), AP50:95 (62.8%), and AR50:95 (75.0%) on the coal rock image dataset.
• Reduced model parameters to 2.59M and FLOPs to 6.9G, with a model weight file size of 5.2 MB.
• Demonstrated superior performance compared to other commonly used object detection algorithms.

Conclusions:

• YOLOv8-Coal effectively enhances recognition accuracy and processing speed for coal-rock images.
• The novel combination of modules leads to a more efficient and accurate object detection system.
• This method offers a promising solution for industrial applications requiring robust image recognition.