Mineral sorting processes, including manual sorting, flotation, gravity separation, and magnetic separation, face critical technical challenges such as low efficiency, high error rates, and unstable concentrate grades. Traditional methods often struggle with complex ore textures, overlapping particles, and dynamic environmental conditions, leading to suboptimal industrial outcomes. To address these limitations, we propose an advanced industrial vision detection system based on an enhanced YOLOv10 architecture. We aim to improve feature extraction capabilities, multi-scale fusion efficiency, and detection accuracy in complex ore scenarios, thereby providing a robust technical foundation for intelligent upgrades in mineral processing workflows.

The proposed system introduces four key innovations to the YOLOv10 framework: 1) The C2f module in the backbone network is enhanced. The Bottleneck in the C2f module is replaced with Bottleneck-CloAtt, which embeds an attention mechanism with a dual-branch architecture. This helps the model to comprehensively grasp the distribution of molybdenum ore targets and their relative relationships with the background, thus avoiding misjudgments. 2) The FocalModulation module is used to replace the SPPF module. This module adopts a focal modulation mechanism instead of the traditional self-attention mechanism. It can capture long-range dependencies and contextual information in images, thus significantly increasing the detection rate of small-sized molybdenum ore targets that are difficult to detect in images. It ensures effective processing of X-ray images of molybdenum ores of various sizes and avoids information loss. 3) The Dysample module is introduced as an upsampler to replace the original Upsample module. While ensuring detection performance, it reduces the computational load and processing time, improves the model’s operational speed, and meets the requirements of industrial real-time monitoring for rapid image processing and timely result output. 4) GIoU loss function: The generalized intersection-over-union (GIoU) loss is adopted to refine bounding box regression accuracy, particularly for irregularly shaped ore particles. A high-resolution molybdenum ore image dataset is constructed using X-ray imaging equipment. The improved model, termed YOLOv10s_pro, is trained and validated on this dataset using a transfer learning strategy.

The proposed YOLOv10s_pro achieves outstanding performance. It demonstrates a precision of 96.5%, representing a 4.9 percentage points improvement over the baseline. The recall rate reaches 96.5%, with a 5.2 percentage points increase compared to the baseline. In terms of mAP@50, it achieves 98.6%, showing a 2.0 percentage points boost, and 79.6% at mAP@50:95, marking a 5.2 percentage points improvement. Each improvement strategy has demonstrated its effectiveness. For example, the C2f module enhanced by CloAtt improves the feature extraction ability; the FocalModulation module optimizes sample weight adjustment to handle hard-to-classify samples; the dynamic upsampler Dysample enhances detection performance for small targets and occluded ores; and the GIoU loss function optimizes prediction box regression to improve detection accuracy and stability, optimizing model performance in different aspects. The performance of the YOLOv10s_pro model, which comprehensively applies these improvement strategies, is further enhanced, fully demonstrating the synergy and mutual promotion relationship among the strategies. For instance, the C2f module enhanced by CloAtt and the Focal Modulation module cooperate in feature extraction and sample weight adjustment, while the dynamic upsampler Dysample and the GIoU loss function work together in increasing the feature mAP resolution and optimizing the prediction box regression. Comprehensively applying multiple improvement strategies provides a better solution for ore detection tasks.

We successfully develop an industrial vision detection system that addresses long-standing bottlenecks in mineral sorting processes. By innovatively integrating attention mechanisms, dynamic modulation, deformable sampling, and advanced loss function into the YOLOv10 framework, the proposed system achieves state-of-the-art detection accuracy and robustness. The experimental results validate its potential to enable real-time, high-precision ore sorting, thereby reducing manual intervention, stabilizing concentrate quality, and optimizing resource utilization. Future work will focus on deploying the model in industrial production lines and extending its application to multi-modal ore analysis.