Volume 173

The rising resource demands in emerging economies have intensified resource nationalism in mineral-rich countries, necessitating more efficient mineral processing technologies for declining ore grades. This study presents BGF-YOLO, a novel deep learning model enhanced from YOLOv8, designed to optimize mineral beneficiation by accurately identifying mineral species and grain sizes using hyperspectral imaging. The system utilizes hyperspectral data spanning 66 spectral bands (400–1000 nm) and processes large datasets through advanced feature fusion and attention mechanisms. BGF-YOLO integrates a Generalized Feature Pyramid Network (GFPN), Dual-Level Routing Attention (DLRA), and an additional detection head to improve multi-scale feature detection and reduce redundant information. Evaluated on a dataset of 4,975 samples across five mineral classes, the model achieved an overall accuracy of 91.9%, with Galena and Hematite large particles attaining 94.9% and 100.0% accuracy, respectively. Comparative analysis showed that BGF-YOLO outperforms the baseline YOLOv8 by approximately 5% in accuracy. These results demonstrate the potential of combining hyperspectral imaging with advanced deep learning architectures to enhance the precision and efficiency of mineral classification and grain size determination in beneficiation processes.

Show more

This paper proposes a unified computational framework, which ensures the output quality of large language models in writing education through three major modules: style transformation, dependency detection, and adversarial intervention. The style conversion module adopts the Transformer model with a dual-encoder architecture to transcribe students' texts into academic or news styles while retaining the original meaning. The dependency detection module reconstructs sentence-level grammatical relations and text-level argumentation structures based on the two-layer graph attention network (GAT). The adversarial intervention module simulates typical student errors through controlled perturbations such as synonym replacement and clause recombination to evaluate the robustness of the model. Experiments show that the academic accuracy rate of the style conversion module reaches 91.8%, the news accuracy rate reaches 89.5%, and the average score of UEBL is 28.6. In the case of adversarial perturbation, the style accuracy rate decreased by only 3.2 percentage points. The syntactic annotation accuracy (LAS) of the GAT parser on the original data was 87.5%, the text F1 value reached 78.3%, and the losses under adversarial interference were controlled at 4.8% (LAS) and 5.3% (F1) respectively. These findings confirm that adversarial training can significantly increase the model's resistance to writing errors. This framework provides educators with practical tools to ensure writing style standardization, structural consistency, and the ability to resist error feedback, laying the foundation for building a reliable AI-assisted writing teaching system.

Show more

Existing algorithms for unmanned aerial vehicle (UAV) image object detection often face challenges such as low detection accuracy for small objects and missed detections of multi-scale objects. To address these issues, this paper proposes a UAV image object detection algorithm that integrates a channel attention mechanism with parallel-structured dilated convolution feature fusion. To enhance the algorithm’s feature representation capabilities in terms of channel attention and receptive field, the ResNet50 backbone is redesigned by incorporating the Squeeze-and-Excitation Network (SENet) and a Parallel-Structured Dilated Convolution Feature Fusion Network (PSDCFFN). Additionally, Region of Interest (ROI) Align is employed, and the Region Proposal Network (RPN) anchor sizes are optimized using K-Means clustering to minimize coordinate deviations during object regression. Experimental results demonstrate that the proposed algorithm significantly improves object detection accuracy in UAV images. On the RSOD-Dataset and a custom UAV image dataset, the mean Average Precision (mAP) reaches 92.52% and 98.07%, respectively.

Show more

To address the challenge of missed detections of long-distance targets in autonomous driving, this study proposes an enhanced 3D object detection model based on the CenterFusion framework, integrating camera and millimeter-wave radar data. An early fusion strategy is employed to project radar data onto the image plane, combining it with image data to form a multi-channel input, thereby enhancing the model’s robustness against interference. Additionally, an attention mechanism is incorporated post-feature fusion to prioritize the extraction of critical information from the fused feature map, significantly improving detection accuracy. The loss function is optimized to mitigate the imbalance between positive and negative samples. Comparative and ablation experiments conducted on the nuScenes dataset demonstrate that the proposed model achieves a 1.5% improvement in average detection accuracy and a 2.1% increase in nuScenes Detection Score (NDS) compared to the baseline CenterFusion model, effectively enhancing long-distance target detection capabilities.

Show more

In this paper, by systematically evaluating the performance of multiple machine learning models in the task of advertisement click prediction, it is found that the XGBoost algorithm exhibits the best prediction potential by virtue of its integrated learning advantages. In order to further improve the model performance, the Sparrow Search Algorithm (SSA) is innovatively introduced to intelligently search for the key hyperparameters of XGBoost, and the SSA-XGBoost fusion model is constructed. The experimental results show that the optimized model achieves significant breakthroughs in classification performance: the accuracy rate reaches 0.87, which is 18.1% higher than that of the basic XGBoost; the recall rate is synchronously increased to 0.87, while the precision rate achieves a leapfrog growth to reach the excellent level of 0.887, which is 21.3% higher than that of the unoptimized model (0.731). These performance improvements have special value in the dimension of false alarm rate reduction - when the model accuracy rate is increased by 21.3%, it means that about 50,000 invalid placements can be reduced in a million-volume ad exposure scenario, and this accuracy improvement not only verifies the effectiveness of the sparrow search algorithm in parameter optimization, but also highlights the practical business value brought by the algorithm improvement. This improvement in accuracy not only verifies the effectiveness of the algorithm in terms of parameter optimization, but also highlights the practical commercial value of the algorithm improvement. From the perspective of feature engineering, SSA successfully solves the efficiency bottleneck of traditional grid search in high-dimensional parameter space through the strategy of combining global search and local optimization, so that key hyperparameters such as the tree structure parameters and learning rate of XGBoost reach a more optimal configuration, which effectively mitigates the risk of overfitting while maintaining the model's stronger generalization ability (19.2% improvement in F1-score) (34% reduction in cross-validation variance). The intelligent prediction model constructed in this study is of great practical significance to the field of digital marketing: through high-precision click prediction, advertisers can accurately identify potential user groups, and reduce the cost of ineffective advertisement exposure while improving the conversion efficiency. This data-driven decision support can not only optimize the advertising budget allocation strategy, but also promote the programmatic advertising delivery system to evolve in the direction of intelligence, and provide technical support for enterprises to build core advantages in digital marketing competition.

Show more

Rural agricultural economies need tailored strategies and intensive cultivation for sustainable development. In China, with limited arable land and a push for agricultural modernization, choosing the right crop planting strategies is essential to meet people's needs and boost agricultural production and the economy.For Problem 1, we created an integer programming model aiming for stable annual economic returns, with average annual planting revenue as the objective function. Two scenarios were considered: excess crops wasted or sold at half price. Using the simulated annealing algorithm, we found that the average annual revenue is 7,882,002.50 yuan in the first scenario and 8,659,569.25 yuan in the second.For Problem 2, we built a robust optimization model to account for the dynamic nature of the agricultural market, including potential risks from rising costs, falling prices, and declining demand. The model focuses on worst-case scenarios to develop a resilient planting strategy, reducing risks from market volatility. It uses minimum annual planting revenue as the objective function and parameter uncertainty sets to create a plan less sensitive to disturbances and effective under most conditions.For Problem 3, we constructed evaluation indicators to explore the substitutability and complementarity between crops, as well as their sales, planting costs, and prices. A grey relational analysis model was used to assess crop similarity. Prioritizing crops with higher returns and stable prices, we selected 29 crops, including legumes, for a new planting strategy. Compared to the strategy from Problem 2, the new plan has fewer crop rotations, more stable economic returns, and easier field management.This paper summarizes and analyzes the established models, providing a comprehensive evaluation of their advantages and limitations.

Show more

This paper, we propose a new mineral image classification method using YOLOv8 model enhanced by visual attention mechanism. The integration of attention blocks improves the model's ability to focus on relevant features, thereby reducing misclassification and improving accuracy, especially in complex and noisy environments. Experimental results using iron ore with different densities show that the attention-enhanced YOLOv8 outperforms traditional methods in ferrous iron type classification based on density, ash and microfraction. The proposed method significantly improves the efficiency of feature extraction and processing, which provides a promising solution for intelligent ore sorting in industrial applications.

Show more