摘要
针对基于卷积神经网络(CNN)的目标检测算法因未对高层特征语义信息和低层特征细节信息之间的关系进行充分利用而导致的小目标检测率低的问题,提出一种基于卷积神经网络的小目标检测改进算法。通过对稠密连接块进行全连接,将多层特征图的特征语义信息进行融合,在候选区域添加一个具有短捷径连接的卷积与反卷积网络,以加快收敛速度。在PASCAL VOC数据集上的实验结果表明,与目前最好的算法相比,小目标的检测率从78.4%、80.5%提高到了81.6%。
Aiming at the problem that the object detection algorithm based on convolutional neural network(CNN) fails to make full use of the relationship between the semantic information of high-level features and the detailed information of low-level features, which leads to the low detection rate of small objects, an improved object detection algorithm based on convolutional neural network is proposed. The feature semantic information of the multi-layer feature graph is fused by fully connecting the dense block. A convolution and deconvolution network with short-cut connection is added to the region proposals to speed up the convergence. Experimental results on the PASCAL VOC data set show that the detection rate of small objects increased from 78.4% and 80.5% to 81.6% compared with the best current algorithms.
作者
吕方方
陈光喜
刘家畅
胡灵
李翘楚
LYU Fangfang;CHEN Guangxi;LIU Jiachang;HU Ling;LI Qiaochu(School of Computer and Information Security,Guilin University of Electronic Technology,Guilin 541004,China;Department of Computer Science and Technology,Tsinghua University,Beijing 100084,China;School of Geosciences,Chengdu University of Technology,Chengdu 610059,China)
出处
《桂林电子科技大学学报》
2021年第5期368-374,共7页
Journal of Guilin University of Electronic Technology
基金
国家自然科学基金(61972225,61462018)
广东省数学教育软件工程技术研究中心公开基金(LD16124X)
广西学位与研究生教育改革项目(JGY2014060)
桂林电子科技大学研究生教育创新计划(2016XWYJ09)。
关键词
小目标
特征信息
稠密连接块
全连接
short-cut连接
small objects
feature information
densely connection blocks
full connection
short-cut connection
