| 引用本文： |
-
陈鹏宇,王烈,梁钰墁,等.AFL-YOLO:基于 YOLOv8 改进的小目标检测算法[J].电讯技术，2026，66（2）： - . [点击复制]
- CHEN Pengyu,WANG Lie,LIANG Yuman,et al.AFL-YOLO:an Improved Small Object Detection Algorithm Based on YOLOv8[J].，2026，66（2）： - . [点击复制]
|
|
| |
|
|
| 本文已被：浏览 22次 下载 5次 |
码上扫一扫！ |
| AFL-YOLO:基于 YOLOv8 改进的小目标检测算法 |
| 陈鹏宇,王烈,梁钰墁,何广斌,陈洪帅 |
|
|
| (广西大学 计算机与电子信息学院，南宁 530004;广西大学 计算机与电子信息学院，南宁 530005;广西大学 计算机与电子信息学院，南宁 530006;广西大学 计算机与电子信息学院，南宁 530007;广西大学 计算机与电子信息学院，南宁 530008) |
|
| 摘要: |
| 针对小目标检测任务中存在的精度低、漏检和误检等问题,提出了一种基于YOLOv8n的小目标检测算法AFL-YOLO。首先,引入Shape-IoU损失函数,其动态特征适应机制能够更好聚焦于边框的形状与尺度,实现更准确的边框回归。其次,在骨干网络中融入空间深度转换卷积(Space-to-DepthConvolution,SPD-Conv),改善细粒度信息丢失的问题。然后,引入感受野注意力卷积(Receptive-FieldAttentionConvolution,RFAConv),构建C2F_RFAConv模块,增强模型对全局上下文信息特征的学习能力。最后,优化检测层,提高检测精度,降低模型参数量。实验表明,AFL-YOLO在VisDrone2019数据集上相比YOLOv8n,mAP@0.5、mAP@0.5:0.95、精度和召回率分别提升了5.3%、3.6%、4.4%、4%,同时参数量减少了20.5%。此外,还在TinyPerson数据集上进行了泛化对比实验,证明提出的AFLYOLO算法在保证模型轻量化的同时,显著提高了对小目标物体的检测精度。 |
| 关键词: 小目标检测 YOLOv8 损失函数 空间深度转换卷积 感受野注意力卷积 |
| DOI：10.20079/j.issn.1001-893x.241024002 |
|
| 基金项目:广西重点研发计划(桂科 AB24010033) |
|
| AFL-YOLO:an Improved Small Object Detection Algorithm Based on YOLOv8 |
| CHEN Pengyu,WANG Lie,LIANG Yuman,HE Guangbin,CHEN Hongshuai |
| (School of Computer and Electronic Information,Guangxi University,Nanning 530004,China;School of Computer and Electronic Information,Guangxi University,Nanning 530005,China;School of Computer and Electronic Information,Guangxi University,Nanning 530006,China;School of Computer and Electronic Information,Guangxi University,Nanning 530007,China;School of Computer and Electronic Information,Guangxi University,Nanning 530008,China) |
| Abstract: |
| For the problems such as low accuracy,missed detections,and false positives in small object detection tasks, an algorithm named AFL-YOLO based on YOLOv8n is proposed. The algorithm is characterized by several enhancements introduced to improve its performance. Firstly,the Shape-IoU loss function is incorporated,whereby the dynamic feature adaptation mechanism allows for a more precise focus on the shape and scale of bounding boxes,leading to improved regression of these boxes. Secondly,spaceto-depth convolution(SPD-Conv) is embedded within the backbone network,mitigating the problem of finegrain information loss. Thirdly, the receptive-field attention convolution ( RFAConv ), along with the construction of the C2F_RFAConv module,is introduced to strengthen the model’s capability in learning global contextual feature information. Finally, the detection layer undergoes optimization to enhance detection accuracy while reducing the model’s parameter count. Experimental results demonstrate that, compared with YOLOv8n,AFL-YOLO achieves improvements of 5. 3% ,3. 6% ,4. 4% ,and 4% in mAP@ 0. 5,mAP@ 0. 5:0. 95,precision,and recall,respectively,on the VisDrone2019 dataset,with a concurrent reduction of 20. 5% in parameters. Additionally,generalization comparison experiments on the TinyPerson dataset prove that the proposed AFL-YOL algorithm effectively enhances the accuracy of detecting small objects while ensuring the lightweight nature of the model. |
| Key words: small object detection YOLOv8 loss function space-to-depth convolution receptive-field attention convolution |
|
|
|
|
|
《电讯技术》编辑部 