文献类型：期刊文献

中文题名：基于改进SSD的鲁棒小目标检测算法

英文题名：Robust small target detection algorithm based on improved SSD

作者：秦振[1,2];李学伟[1,2];刘宏哲[1,2]

第一作者：秦振

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2023

卷号：55

期号：4

起止页码：59-66

中文期刊名：东北师大学报(自然科学版)

外文期刊名：Journal of Northeast Normal University(Natural Science Edition)

收录：CSTPCD;;北大核心:【北大核心2020】；

基金：国家自然科学基金资助项目(62171042,62102033,62006020);北京市重点科技项目(KZ202211417048);北京市属高等学校高水平科研创新团队建设支持计划项目(BPHR20220121);北京市自然科学基金资助项目(4232026).

语种：中文

中文关键词：目标检测;小目标;特征融合;特征增强;非局部关系

外文关键词：object detection;small objects;feature fusion;feature enhancement;non-local relationship

摘要：为提高各种场景中目标检测性能,解决小目标检测困难问题,提高小目标的检测精度,提出一种基于新的特征融合方法与特征增强模块改进的SSD鲁棒小目标检测算法.通过将相邻特征进行融合来有效利用上下文信息,而后通过不同尺寸特征之间的非局部关系来增强特征映射,从而提高对小目标的检测精度,与Faster R-CNN、SSD、SD-SSD、FD-SSD等目标检测模型在PASCAL VOC、KITTI公开数据集对比以验证其小目标检测性能.实验结果表明:所提出方法的A mAP值在两个数据集上达到了81.3%与80.6%,其中针对小目标的检测性能提升尤为显著,在一定程度上解决了小目标检测难的问题,对于自动驾驶等应用有着现实意义.
In order to improve the target detection performance in various scenes,solve the difficult problem of small target detection and improve the detection accuracy of small targets,a robust small target detection algorithm for SSD based on a new feature fusion method with improved feature enhancement module is proposed to effectively utilize contextual information by fusing adjacent features,and then enhance the feature mapping by non-local relationships between features of different sizes to improve the detection accuracy of small the proposed method is compared with Faster R-CNN,SSD,SD-SSD,FD-SSD and other target detection models in PASCAL VOC and KITTI public datasets to verify its small target detection performance.The experimental results show that the mAP values of the proposed method reach 81.3%and 80.6%in the two datasets,and the detection performance for small targets is particularly improved,which solves the problem of difficult detection of small targets to a certain extent and has practical significance for applications such as autonomous driving.