详细信息
文献类型:期刊文献
中文题名:面向社群图像的显著区域检测方法
英文题名:Salient region detection for social images
作者:梁晔[1,2];于剑[2]
第一作者:梁晔
机构:[1]北京联合大学机器人学院;[2]北京交通大学计算机与信息技术学院
第一机构:北京联合大学机器人学院
年份:2018
卷号:13
期号:2
起止页码:174-181
中文期刊名:智能系统学报
外文期刊名:CAAI Transactions on Intelligent Systems
收录:CSTPCD;;北大核心:【北大核心2017】;CSCD:【CSCD2017_2018】;
基金:北京市自然科学基金项目(4182022);北京联合大学2017年度人才强校百杰计划项目(BPHR2017CZ10);"十三五"时期北京市属高校高水平教师队伍建设支持计划项目(IDHT20170511);国家科技支撑计划项目(2015BAH55F03)
语种:中文
中文关键词:显著性;显著区域;社群图像;深度学习;标签
外文关键词:saliency; salient region; social images; deep learning; tag
摘要:网络技术和社交网站的发展带来了社群图像的飞速增长。海量的社群图像成为了非常重要的图像类型。本文关注社群图像的显著区域检测问题,提出基于深度特征的显著区域检测方法。针对社群图像带有标签的特点,在系统框架中,本文采取两条提取线:基于CNN特征的显著性计算和基于标签的语义计算,二者的结果进行融合。最后,通过全连接的条件随机场模型对融合的显著图进行空间一致性优化。此外,为了验证面向社群图像的显著区域检测方法的性能,针对目前没有面向社群图像的带有标签信息的显著性数据集,基于NUS-WIDE数据集,本文构建了一个图像结构丰富的社群图像数据集。大量的实验证明了所提方法的有效性。
The development of network technology and social website has brought about the rapid growth of social images.Massive social images have become a very important image type.This paper focuses on the detection problem of salient region for social images,a method for detecting salient region and based on depth features was proposed.By considering the feature that the social image is attached with tag,in the framework of the system,the paper used two extraction lines:the saliency computing based on CNN features and the semantic computing based on tag,the results of both parts were fused.Finally,saliency maps were optimized by a fully connected conditional random field model for the spatial consistency.In addition,for verifying the performances of the saliency region detection method orienting social image,in view of the lack of saliency dataset with tags for social images,on basis of NUS-WIDE dataset,the paper constructed a social image dataset with rich image structures.Extensive experiments demonstrated the effectiveness of the proposed method.
参考文献:
正在载入数据...