基于YOLOv8s的航拍小目标检测算法轻量化与改进  

Lightweight and Improvement of Aerial Photography Small Target Detection Algorithm Based on YOLOv8s

作　　者：吴建祥 李晓飞[1] WU Jianxiang;LI Xiaofei(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《软件导刊》2024年第11期193-199,共7页Software Guide

摘　　要：小目标检测一直是目标检测领域的热点,而航拍小目标图像具有尺度变化大、视角和姿态变化多样等特点,容易造成漏检、误检。针对这些问题,提出基于YOLOv8s改进的轻量化小目标检测算法。首先,在Backbone部分使用CESE_C2f模块,利用轻量化注意力机制提升检测效果并降低参数量;其次,根据BiFPN思想设计新的网络结构,充分利用已有网络信息加强深层与浅层特征信息的融合;最后,使用新的轻量化检测头LW_Detect(Lightweight_Detect)大幅度降低运算量与参数量,以满足实时性要求。实验结果表明,所提算法在Visdrone数据集上的mAP50和mAP50:95分别达到41.4%和24.9%,相较原始YOLOv8s算法精度提高了1.6%和1.3%,参数量和运算量降低了36.8%和42.8%。Small target detection has always been a hot topic in the field of object detection,and aerial images of small targets have the charac⁃teristics of large scale changes,diverse perspectives and poses,which can easily lead to missed or false detections.To address these issues,a lightweight small object detection algorithm based on YOLOv8s improvement is proposed.Firstly,the CESE-C2f module is used in the Back⁃bone section to enhance detection performance and reduce parameter count by utilizing a lightweight attention mechanism;Secondly,design a new network structure based on the BiFPN concept,fully utilizing existing network information to enhance the fusion of deep and shallow fea⁃ture information;Finally,the new lightweight detection head LW-Detect is used to significantly reduce the computational and parameter re⁃quirements to meet real-time requirements.The experimental results showed that the mAP50 and mAP50:95 of the proposed algorithm on the Visdrone dataset reached 41.4%and 24.9%,respectively.Compared with the original YOLOv8s algorithm,the accuracy was improved by 1.6%and 1.3%,and the parameter and computational complexity were reduced by 36.8%and 42.8%,respectively.

关 键 词：YOLOv8s 小目标 轻量化 BiFPN 目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]