| 黄成,林生佐,吴建东.基于YOLOv8n的航拍地面目标检测改进算法[J].高技术通讯(中文),2026,36(6):593~601 |
| 基于YOLOv8n的航拍地面目标检测改进算法 |
| Improved algorithm for aerial ground target detection based on YOLOv8n |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 06. 005 |
| 中文关键词: YOLOv8n； 目标检测； 航拍 |
| 英文关键词: YOLOv8n, object detection, aerial photography |
| 基金项目: |
| 作者 | 单位 | | 黄成 | （广东环境保护工程职业学院实训中心佛山 528216） | | 林生佐 | | | 吴建东 | |
|
| 摘要点击次数: 85 |
| 全文下载次数: 66 |
| 中文摘要: |
| 针对传统目标检测算法在航拍图像小目标检测中效果不佳的情况，本文基于单次目标检测器第8代最小（you only look once version 8 nano，YOLOv8n)模型进行了改进。在C2f(Conv 2 fusion)模块中引入滑动窗口变换器(Swin Transformer)模块和多尺度特征融合技术，提高了模块对不同尺寸目标的特征提取能力；提出了平方距离交并比(squared distance intersection over union，SDIoU)损失函数，通过对角距离平方与平均面积比的优化策略，使定位预测更为准确。设计了多维小目标检测头，从多个维度对目标进行检测，大幅减少小目标的漏检。在VisDrone 2019数据集上进行实验，改进后的模型在关键性能指标mAP@50和mAP@50∶95上，较原模型分别提高了5.8%和3.7%。 |
| 英文摘要: |
| To address the poor performance of traditional object detection algorithms in small object detection for aerial images—attributed to challenges like low target resolution and complex backgrounds—this study proposes three key improvements based on the YOLOv8n (you only look once version 8 nano) model. First, the Swin Transformer module and multi-scale feature fusion technology are embedded into the original C2f (conv 2 fusion) module, forming an enhanced feature extraction unit that strengthens the model’s capability to capture fine-grained features of objects across different scales, particularly small ones. Second, a square distance intersection over union (SDIoU) loss function is developed; by integrating an optimization strategy that combines squared diagonal distance and average area ratio, this function effectively reduces bounding box regression errors, enabling more accurate localization of small targets. Third, a multi-dimensional small object detection head is designed to analyze targets from semantic, spatial, and scale dimensions, which significantly mitigates the missed detection of small objects in aerial scenes. Comprehensive experiments were conducted on VisDrone 2019, a widely recognized benchmark dataset for aerial object detection. The results demonstrate that the improved model achieves mAP@50 of 0.406 and mAP@50∶95 of 0.241, representing increases of 5.8% and 3.7%, respectively, compared to the original YOLOv8n model. This outcome validates that the proposed improvements effectively balance detection accuracy and efficiency for small object detection in practical aerial scenarios. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
