YOLO-MARS: An Enhanced YOLOv8n for Small Object Detection in UAV Aerial Imagery DOI Creative Commons
Guofeng Zhang, Yanfei Peng, Jincheng Li

et al.

Sensors, Journal Year: 2025, Volume and Issue: 25(8), P. 2534 - 2534

Published: April 17, 2025

In unmanned aerial vehicle (UAV) imagery scenarios, challenges such as small target size, compact distribution, and mutual occlusion often result in missed detections false alarms. To address these challenges, this paper introduces YOLO-MARS, a recognition model that incorporates multi-level attention residual mechanism. Firstly, an ERAC module is designed to enhance the ability capture targets by expanding feature perception range, incorporating channel weight allocation strategies strengthen extraction capability for introducing connection mechanism improve gradient propagation stability. Secondly, PD-ASPP structure proposed, utilizing parallel paths differentiated depthwise separable convolutions reduce computational redundancy, thereby enabling effective identification of at various scales under complex backgrounds. Thirdly, multi-scale SGCS-FPN fusion architecture adding shallow guidance branch establish cross-level semantic associations, effectively addressing issue loss deep networks. Finally, dynamic WIoU evaluation function implemented, constructing adaptive penalty terms based on spatial distribution characteristics predicted ground-truth bounding boxes, optimizing boundary localization accuracy densely packed from UAV viewpoint. Experiments conducted VisDrone2019 dataset demonstrate YOLO-MARS method achieves 40.9% 23.4% mAP50 mAP50:95 metrics, respectively, representing improvements 8.1% 4.3% detection compared benchmark YOLOv8n, thus demonstrating its advantages detection.

Language: Английский

GC-Faster RCNN: The Object Detection Algorithm for Agricultural Pests Based on Improved Hybrid Attention Mechanism DOI Creative Commons

Bolun Guan,

Yaqian Wu,

Jingbo Zhu

et al.

Plants, Journal Year: 2025, Volume and Issue: 14(7), P. 1106 - 1106

Published: April 2, 2025

Pest infestations remain a critical threat to global agriculture, significantly compromising crop yield and quality. While accurate pest detection forms the foundation of precision management, current approaches face two primary challenges: (1) scarcity comprehensive multi-scale, multi-category datasets (2) performance limitations in models caused by substantial target scale variations high inter-class morphological similarity. To address these issues, we present three key contributions: First, introduce Insect25-a novel agricultural dataset containing 25 distinct categories, comprising 18,349 high-resolution images. This specifically addresses diversity through multi-resolution acquisition protocols, enriching feature distribution for robust model training. Second, propose GC-Faster RCNN, an enhanced framework integrating hybrid attention mechanism that synergistically combines channel-wise correlations spatial dependencies. dual design enables more discriminative extraction, which is particularly effective distinguishing morphologically similar species. Third, implement optimized training strategy featuring cosine annealing scheduler with linear warm-up, accelerating convergence while maintaining stability. Experiments have shown compared original Faster RCNN model, has improved average accuracy mAP0.5 on Insect25 4.5 percentage points, mAP0.75 20.4 mAP0.5:0.95 increased 20.8 recall rate 16.6 points. In addition, experiments also method can reduce interference from multiple scales similarity between improving performance.

Language: Английский

Citations

1
YOLO-MARS: An Enhanced YOLOv8n for Small Object Detection in UAV Aerial Imagery DOI Creative Commons
Guofeng Zhang, Yanfei Peng, Jincheng Li

et al.

Sensors, Journal Year: 2025, Volume and Issue: 25(8), P. 2534 - 2534

Published: April 17, 2025

In unmanned aerial vehicle (UAV) imagery scenarios, challenges such as small target size, compact distribution, and mutual occlusion often result in missed detections false alarms. To address these challenges, this paper introduces YOLO-MARS, a recognition model that incorporates multi-level attention residual mechanism. Firstly, an ERAC module is designed to enhance the ability capture targets by expanding feature perception range, incorporating channel weight allocation strategies strengthen extraction capability for introducing connection mechanism improve gradient propagation stability. Secondly, PD-ASPP structure proposed, utilizing parallel paths differentiated depthwise separable convolutions reduce computational redundancy, thereby enabling effective identification of at various scales under complex backgrounds. Thirdly, multi-scale SGCS-FPN fusion architecture adding shallow guidance branch establish cross-level semantic associations, effectively addressing issue loss deep networks. Finally, dynamic WIoU evaluation function implemented, constructing adaptive penalty terms based on spatial distribution characteristics predicted ground-truth bounding boxes, optimizing boundary localization accuracy densely packed from UAV viewpoint. Experiments conducted VisDrone2019 dataset demonstrate YOLO-MARS method achieves 40.9% 23.4% mAP50 mAP50:95 metrics, respectively, representing improvements 8.1% 4.3% detection compared benchmark YOLOv8n, thus demonstrating its advantages detection.

Language: Английский

Citations

1