针对目标检测任务中小目标特征信息不足、检测率较低,且错、漏检率较高等缺点,提出一种基于多尺度特征融合以及混合注意力机制的 Tr-SSD算法.首先,使用 Resnet50 残差网络作为 SSD算法的骨干网络,增强 SSD 算法的特征提取能力;其次,设计了一种混合注意力机制并将其应用于网络的中尺度特征图中以增强特征图中的有效信息,并建立信息间的远距离依赖;最后,使用以 Transformer为核心的网络层与替换骨干网络后的 SSD算法形成FPN结构,融合不同尺度的特征信息,以更准确地对小目标进行定位.实验结果表明:Tr-SSD 算法在 PASCAL VOC数据集、HRSID数据集和RSOD遥感数据集上检测的mAP值分别达到 81.9%、87.5%和 88.4%,比SSD算法分别提高了 4.7 百分点、6.8 百分点和 9.2 百分点,且检测速度均满足实时检测的要求.
To address to the low feature information,low detection rates,and high false rate and missing rate in the target detection task,a Tr-SSD algorithm based on multiscale feature fusion and a hybrid attention mechanism was proposed.Firstly,a Resnet50 residual network was utilized as the backbone network for the SSD algorithm to en-hance its feature extraction capabilities.Secondly,a hybrid attention mechanism was designed and applied to the mid-scale feature maps of the network to enhance effective information within the feature maps and establish long-range dependencies between pieces of information.Finally,a FPN(feature pyramid network)structure was formed by using network layers centered around the Transformer instead of the original backbone network in the SSD algo-rithm,which fused feature information of different scales to more accurately locate small targets.Experimental re-sults showed that the Tr-SSD algorithm achieved mAP values of 81.9%,87.5%,and 88.4%on the PASCAL VOC dataset,HRSID dataset,and RSOD remote sensing dataset,respectively.This represented an improvement of 4.7 percentage points,6.8 percentage points,and 9.2 percentage points compared to the original SSD algorithm.Mo-reover,the detection speed could meet the requirements for real-time detection.
魏明军;王镆涵;刘亚志;李辉
华北理工大学 人工智能学院,河北 唐山 063210||华北理工大学 河北省工业智能感知重点实验室,河北 唐山 063210华北理工大学 人工智能学院,河北 唐山 063210
计算机与自动化
小目标检测;注意力机制;特征融合;深度学习;实时检测
small target detection;attention mechanism;feature fusion;deep learning;real-time detection
《郑州大学学报（工学版）》 2024 (003)
72-79 / 8
科技部重点研发项目(2017YFE0135700);河北省高等学校科学技术研究项目(ZD2022102)
10.13705/j.issn.1671-6833.2024.03.001
评论