基于复合缩放的动态航拍小目标检测算法  

Dynamic Aerial Small Target Detection Algorithm Based on Compound Zoom Scaling

作　　者：蒋源 朱高峰 朱凤华[3] 熊刚[3] Jiang Yuan;Zhu Gaofeng;Zhu Fenghua;Xiong Gang(Huzhou Vocational&Technical College,Huzhou 313000,China;Shandong Jiaotong University,Jinan 250300,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]湖州职业技术学院,浙江湖州313000 [2]山东交通学院,济南250300 [3]中国科学院自动化研究所,北京100190

出　　处：《航空兵器》2025年第2期104-112,共9页Aero Weaponry

基　　金：国家自然科学基金项目(U1909204);江西省自然科学基金项目(20232ABC03A07)。

摘　　要：无人机结合计算机视觉技术成为新的信息获取手段,广泛应用于各个领域,但无人机多角度成像过程中,目标像素占比小,受背景干扰严重,存在目标漏检误检等现象。针对此问题,提出一种小目标检测算法。首先,替换更高效的主干网络,应用复合缩放的方式来探索网络深度、宽度以及图像分辨率三者间的平衡点;结合C2f层级链接的优势,应用注意力机制来有效提取不同尺度、方向和形状目标的上下文细节信息,并通过并行网络实现小目标位置信息的交互建模。其次,针对小目标像素利用率低的特点,设计DTADH模块,构建共享特征交互模块,结合任务对齐预测器进行目标的分类以及定位分配,经注意力机制动态计算任务特征来进行任务分解,有效减少了参数量,提高整体性能。在无人机航拍数据集VisDrone2019上进行实验,结果显示算法的mAP提升了2.1%,FLOPs减少了32.5%。计算复杂度下降,具有更好的检测效果。Unmanned aerial vehicles(UAVs)equipped with computer vision technology have emerged as a po-werful tool for information acquisition and are widely applied across various fields.However,during multi-angle imaging,UAVs often encounter challenges such as a low target pixel ratio and significant background interference,leading to missed detection and false detection.To address these issues,this paper proposes a novel small-object detection algorithm.Firstly,a more efficient backbone network is introduced,and a composite scaling method is employed to optimize the balance among network depth,width,and image resolution.Additionally,an attention mechanism is integ-rated to effectively capture contextual details of targets with varying scales,orientations,and shapes by leveraging the hierarchical connections of the C2f module,and parallel network is further utilized to enhance interactive modeling of small-target positional information.Secondly,to mitigate the issue of low pixel utilization ratio in small-target detection,a DTADH module is designed,and a shared feature interaction module is constructed.This module,coupled with a task alignment predictor,facilitates both target classification and localization allocation,and the task decomposition is performed by dynamically computing task features through an attention mechanism,thereby reducing the number of parameters effectively and enhancing overall performance.Experiments conducted on the VisDrone2019 UAV aerial ima-gery dataset demonstrate that the proposed algorithm improves mAP by 2.1%,reduces FLOP s by 32.5%,and decreases computational complexity,resulting in superior detection performance.

关 键 词：深度学习 目标检测 注意力机制 计算机视觉 复合缩放 航拍 无人机 

分 类 号：TJ760[兵器科学与技术—武器系统与运用工程]