YOLOv8-MD: A highly accurate algorithm for detecting ore belt demarcation points

Abstract

Shaking tables are essential gravity separation equipment widely used in the mineral processing industry, particularly for fine particle concentration. The rapid development of deep learning techniques has brought notable advances in detecting ore belt demarcation points on shaking tables. YOLOv8 improves detection efficiency through architectural refinement and the integration of advanced components. However, it faces limitations when dealing with complex scenes, such as insufficient extraction of low-level features and inadequate channel representation in high-level feature maps. To address these issues, four novel modules were designed: the Dual-Dimensional Feature Attention (DDFA) module and the Deep Cross-Modality Enhanced Attention (DCMEA) module, intended to enhance shallow spatial-channel information and deep semantic representations, respectively. Unlike conventional attention mechanisms such as SE and CBAM, which process channel or spatial cues independently, DDFA strengthens both spatial localization and channel selectivity in low-level features, while DCMEA incorporates multi-scale context and modality-aware attention to enhance high-level semantic understanding. Additionally, the Multiscale Feature Alignment (MFA) module was developed to improve the perception of fine-grained features, and the lightweight C2f-Star module was introduced to reduce parameter size and enable efficient deployment on edge devices. Experimental results indicate that the proposed model achieves superior performance over five mainstream object detection algorithms, with a 7.13% increase in precision and a 6.02% gain in recall compared to the original YOLOv8. Industrial tests show an 83.5% reduction in variance and a 2.3% increase in ore grade, confirming the model’s practical value.