基于分步多尺度特征融合的SSD目标检测算法
SSD Target Detection Algorithm Based on Stepwise Multi-scale Feature fusion
蒋帅 1薛波1
作者信息
摘要
为解决SSD(single shot multibox detector)目标检测算法由于浅层特征表征能力不强,导致对小目标检测精度较低的问题,提出了一种基于分步多尺度特征融合的SSD目标检测算法.为增加SSD模型浅层特征包含的细节信息和语义信息,在模型的低层特征部分引入了两个特征层;只对模型低层的两个特征图进行反卷积操作,并且分两步将低层三个不同尺度的特征图进行特征融合,不仅提高了模型浅层特征的表征能力,而且减少了算法运行过程中的计算量.实验结果表明,在PASCAL VOC2007数据集上,改进后的算法小目标类别的AP值得到了大幅度提高,mAP值比SSD算法提高了3.6%,算法的检测速度也满足实时性要求.
Abstract
In order to solve the problem that SSD(single shot multibox detector)target detection algorithm has poor detection precision for small targets due to weak shallow feature representation ability,an SSD target detection algorithm based on stepwise multi-scale feature fusion is proposed.In order to increase the details and semantic information contained in the shallow features of SSD model,two feature layer are introduced in the low-level feature part of SSD model.Only the two feature maps of the lower layer of the model are deconvolutioned,and the feature maps of three different scales of the lower layer are fused in two steps,which not only improves the representation ability of the shallow feature of the model,but also reduces the calculation amount in the running process of the algorithm.Experimental results show that,on PASCAL VOC2007 dataset,the AP value of small target category is greatly improved by the improved algorithm,and the mAP value is 3.6%higher than that of SSD algorithm.The detection speed of the algorithm also meets the real-time requirement.
关键词
目标检测/SSD/反卷积/分步多尺度特征融合Key words
target detection/single shot multibox detector/deconvolution/stepwise multi-scale feature fusion引用本文复制引用
基金项目
江苏理工学院研究生实践创新计划(XSJCX21_27)
出版年
2024
国家科技期刊平台
NSTL
万方数据