摘要
针对当前基于深度学习的目标检测算法采取的特征图融合方式存在缺陷,算法普遍不能很好地应对尺度变化等问题,提出一种跨深度卷积特征增强的目标检测算法CDC-YOLO。对YOLOv3算法进行改进,针对多尺度预测层各自的特点采用与之适应的特征增强模块,采用多通道的跨深度的卷积核并结合空洞卷积并行地提取特征,最终级联起来。该模块能充分利用多尺度多深度特征,形成统一的多尺度特征表达。在VOC2007test上的实验结果表明,提出算法的mAP (均值平均精度)高达82.33%,比原始YOLOv3提升了约2%,且对尺度变化大的物体鲁棒性更强。
Aiming at the problems that cascading feature maps directly for object detection based on deep learning is defective,and that algorithms generally cannot handle scale variation very well,a cross-depth convolutional feature enhancement algorithm for object detection called CDC-YOLO was proposed.The YOLOv3 algorithm was improved,and different feature enhancement modules were used for different prediction layers according to the characteristics of the multi-scale prediction layer.Multi-channel convolution with different depths was used and combined with dilated convolution to extract features in parallel,and they were concatenated.The module makes good use of multi-scale and multi-depth features to form a uniform multi-scale feature representation.Experimental results on the VOC2007test show that the CDC-YOLO achieves 82.33%mAP(mean average precision),which is improved by about 2%compared with the original YOLOv3.CDC-YOLO is more robust to the scale of object instances that varies in a wide range.
作者
王若霄
徐智勇
张建林
WANG Ruo-xiao;XU Zhi-yong;ZHANG Jian-lin(Institute of Optics and Electronics,Chinese Academy of Sciences,Chengdu 610209,China;School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《计算机工程与设计》
北大核心
2020年第7期1864-1871,共8页
Computer Engineering and Design
基金
国家863高技术研究发展计划基金项目(G158207)。
关键词
深度学习
目标检测
卷积神经网络
特征增强
多尺度特征
deep learning
object detection
convolutional neural network
feature enhancement
multi-scale features
