YOLOv8-POS: a lightweight model for coal-rock image recognition

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Coal-rock image recognition faces challenges like false detections due to image quality and occlusions.
• Existing models often have high complexity, limiting practical deployment in industrial settings.

Purpose of the Study:

• To introduce YOLOv8-POS, a novel approach for accurate and efficient coal-rock image recognition.
• To reduce model complexity and computational demands while enhancing detection performance.

Main Methods:

• Developed a C2f-PConv module combining C2f and partial convolution (PConv) for selective channel processing.
• Incorporated an Overlapping Spatial Reduction Attention module to optimize spatial feature fusion.
• Implemented a slim-neck design with lightweight modules to reduce computational and storage requirements.

Main Results:

• YOLOv8-POS achieved AP50 of 77.1% and AP50:95 of 63.6% on coal-rock datasets.
• Significantly reduced model parameters to 2.60 M and FLOPS to 6.4 G.
• Demonstrated superior performance compared to other prominent algorithms in comparative evaluations.

Conclusions:

• YOLOv8-POS offers a highly accurate and efficient solution for coal-rock image recognition.
• The model's reduced complexity enhances its practical applicability in real-world scenarios.
• This refined approach presents a significant advantage for industrial deployments.