To address the inefficiencies and inaccuracies of traditional ore grade identification methods in complex mining environments, and the challenge of balancing accuracy and speed on edge devices, this paper proposes a lightweight, high-precision, and high-speed detection model named GOG-RT-DETR. Built on the RT-DETR framework, the model incorporates a Faster-Rep-EMA module in the backbone network to reduce computational redundancy and enhance feature extraction. Additionally, a BiFPN-GLSA module replaces the CCFM module in the Neck network, improving feature fusion between the backbone and Neck networks, thus strengthening the model’s ability to capture both global and local spatial features. A Wise-Inner-Shape-IoU loss function is introduced to optimize the bounding box regression, accelerating convergence and improving localization accuracy. The model is evaluated on a custom-built graphite ore dataset with simulated data augmentation. Experimental results show that, compared to the baseline model, the mAP and FPS of GOG-RT-DETR are improved by 2.5% and 8.2%, with a 26.0% reduction in model parameters and a 23.37% reduction in FLOPs. This model enhances detection accuracy and reduces computational complexity, offering an efficient solution for ore grade detection in industrial applications.
In electronic manufacturing, efficient detection of printed circuit board (PCB) surface defects is essential for reducing rework rates and minimizing material waste, thereby supporting sustainable manufacturing. To address the challenge that existing methods struggle to balance detection accuracy and real-time performance in complex industrial environments, this paper proposes a lightweight and high-performance PCB surface defect detection model, termed SDD-RT-DETR. Built upon Real-Time Detection Transformer (RT-DETR), the proposed model introduces a Faster-Block backbone to improve feature extraction efficiency, replaces the original feature fusion module with HS-FPN to enhance multi-scale representation, and employs the Wise-Focaler-MPDIoU loss to optimize bounding box regression. Experiments conducted on an expanded PCB defect dataset containing 3403 images show that SDD-RT-DETR achieves improvements of 2.3% in mAP and 3.6% in inference speed over the baseline, while reducing parameters by 5.04 M and FLOPs by 12.7 G. These results demonstrate that the proposed method effectively balances accuracy, efficiency, and computational cost, offering a practical solution for low-energy and sustainable intelligent electronic manufacturing systems.
Vision-based sensing is a practical approach for real-time ore grade recognition in intelligent mineral sorting. However, graphite ore images captured in industrial environments are often affected by illumination variation, background interference, surface oxidation, and subtle visual differences among grades, limiting the reliability of existing models on edge devices. To address these issues, this study proposes OGI-RT-DETR, a lightweight real-time end-to-end vision-sensing model for graphite ore grade recognition. Based on RT-DETR-r18, a PConv-Rep backbone enhancement module is designed to reduce redundant computation and improve the extraction of ore texture, edge, and irregular shape features. A CFPT feature fusion module is constructed to strengthen multi-scale semantic interaction and spatial fusion under complex backgrounds. A Wise-Focaler-MPDIoU bounding box regression loss is further proposed to improve localization stability and regression accuracy. Experiments were conducted on a self-built augmented graphite ore image dataset collected using an industrial camera system. The proposed OGI-RT-DETR achieved a Precision of 85.6%, Recall of 87.6%, mAP50 of 86.1%, and inference speed of 84.5 FPS. Compared with the baseline RT-DETR-r18, OGI-RT-DETR improved mAP50 by 3.7 percentage points and FPS by approximately 8.2%, while reducing FLOPs and parameters by 31.8% and 38.2%, respectively. These results indicate that OGI-RT-DETR can support accurate, efficient, and lightweight visual sensing for real-time graphite ore sorting.
Real-time instance segmentation of low-grade graphite ores is essential for intelligent mineral-sorting systems, yet existing methods often struggle to balance segmentation accuracy, computational efficiency, and lightweight deployment under irregular boundaries, multi-scale targets, and complex illumination conditions. To address these challenges, this paper proposes LGG-YOLO-Seg, a lightweight and efficient instance segmentation framework based on YOLO11-seg. First, a compact baseline is constructed through depth scaling to reduce redundant computation and model complexity. Then, Linear Deformable Convolution (LDConv) is embedded into the backbone to enhance adaptive feature extraction for irregular ore boundaries. Meanwhile, the feature extraction module is redesigned as C3k2-PKI using multi-kernel and depthwise separable convolutions to improve multi-scale representation capability. In addition, a lightweight Segment-EfficientHead is introduced to reduce computational overhead while maintaining segmentation accuracy. Experiments conducted on a self-constructed graphite ore dataset demonstrate that LGG-YOLO-Seg improves mAP50box and mAP50mask by 0.9% and 0.4%, respectively, compared with YOLO11n-seg while reducing FLOPs, parameters, and model size by 28.8%, 55.6%, and 55.1%, respectively, with an 11.0% increase in inference speed, showing strong potential for real-time industrial sorting and edge deployment.
Accurate identification of low-grade graphite ore is important for improving resource utilization and intelligent mineral sorting. To reduce the computational burden of existing instance segmentation models, this study proposes a lightweight model, RVE-YOLO-seg, based on YOLOv12-seg. GhostConv is introduced for lightweight downsampling, C3k2-RVE is designed to enhance fine-grained feature representation, and Segment-SEAM is employed to strengthen mask-oriented feature extraction. Experiments on a self-constructed dataset of 1978 conveyor-belt images and 19,614 annotated ore instances show that RVE-YOLO-seg achieves 92.7% mAP50 for bounding boxes and 86.3% mAP50 for masks, comparable to YOLOv12n-seg. Meanwhile, the parameter count, FLOPs, and model size are reduced by 56.7%, 24.6%, and 52.5%, respectively, with an inference speed of 80 FPS. Multi-seed experiments and retraining on an external mineral-image dataset further demonstrate stable performance and cross-dataset applicability. These results indicate that RVE-YOLO-seg achieves a favorable accuracy–efficiency trade-off for resource-constrained graphite ore sorting.