| 摘要: |
| 航拍图像中，由于目标较小、特征不明显且分布密集，导致对小目标的检测存在漏检和误检严重等问题。为了解决这些问题，并考虑到无人机设备计算资源的有限性，提出了一种基于YOLOv8n的小目标检测改进算法：LS-YOLO（Light and Scale-YOLO）。首先，为了避免小目标在特征提取和增强过程中被噪声和冗余信息淹没，改进算法去除了主干网络中最后一层特征提取层，再将原P5检测头替换为P2小目标检测头；然后，在主干网络中新设计了轻量化多尺度特征提取模块（Lightweight Multi-scale Convolution，LMC），以更好地提取小目标的多尺度特征，同时减少计算开销并提高改进算法模型的运行效率；最后，新设计共享任务对齐检测头（Shared Task Alignment Detection Head，STA-Head），旨在解决特征在分类和回归任务中不协调的问题，从而提升模型的检测精度，并进一步降低模型的参数量。相比于基线算法YOLOv8n，所设计的算法模型在VisDrone2019数据集上的mAP50提高了8%，达到432%，准确率和召回率分别提高8%和6.7%，分别达到536%和41.4%，参数量下降56.6%，仅有13×106；在RSOD数据集上同样表现良好，展现了良好的泛化能力。 |
| 关键词: 航拍图像 目标检测 小目标 YOLOv8n |
| DOI：10.20079/j.issn.1001-893x.241024003 |
|
| 基金项目:国家自然科学基金资助项目（61831002） |
|
| LS-YOLO：a Small Target Detection Algorithm for Aerial Images Based on Improved YOLOv8n |
| WU Tenghui,DENG Bingguang |
| (School of Communication and Information Engineering，Chongqing University of Posts and Telecommunications，Chongqing 400065，China) |
| Abstract: |
| In aerial images，the detection of small targets is challenged by factors such as small target size，indistinct features，and dense distribution，leading to significant issues with missed and false detections.To address these challenges，and in consideration of the limited computational resources of unmanned aerial vehicles，an improved small target detection algorithm based on YOLOv8n is proposed:LS-YOLO(Light and Scale-YOLO).First，to prevent small targets from being obscured by noise and redundant information during feature extraction and enhancement，the improved algorithm removes the final feature extraction layer in the backbone network and replaces the original P5 detection head with the P2 small target detection head.Second，a lightweight multi-scale feature extraction module，Lightweight Multi-scale Convolution(LMC)，is introduced in the backbone network.This module enhances the extraction of multi-scale features for small targets while reducing computational cost and improving the efficiency of the algorithm.Finally，a new Shared Task Alignment Detection Head(STA-Head) is proposed to address feature misalignment between classification and regression tasks，enhancing detection accuracy and further reducing the model’s parameters.Compared with the baseline YOLOv8n algorithm，the proposed LS-YOLO model achieves an 8% improvement in mean average precision(mAP)50 on the VisDrone2019 dataset，reaching 43.2%.The precision and recall increase by 8% and 6.7%，achieving 53.6% and 41.4%，respectively.The model’s parameters are reduced by 56.6%，totaling only 1.3×106.LS-YOLO also performs well on the RSOD dataset，demonstrating strong generalization ability. |
| Key words: aerial images object detection small target YOLOv8n |
《电讯技术》编辑部 