首页 | 本学科首页   官方微博 | 高级检索  
     

基于深度学习的图像显著对象检测
引用本文:刘春晖,周洋,刘晓琪,唐向宏.基于深度学习的图像显著对象检测[J].光电子.激光,2019,30(1):95-103.
作者姓名:刘春晖  周洋  刘晓琪  唐向宏
作者单位:杭州电子科技大学通信工程学院,杭州,310018;杭州电子科技大学通信工程学院,杭州,310018;杭州电子科技大学通信工程学院,杭州,310018;杭州电子科技大学通信工程学院,杭州,310018
基金项目:国家自然科学基金(61401132,61471348)和浙江省自然科学基金(LY17F020027)资助项目 (杭州电子科技大学 通信工程学院,杭州 310018)
摘    要:显著区域检测可应用在对象识别、图像分割、视 频/图像压缩中,是计算机视觉领域的重要研究主题。然而,基于不 同视觉显著特征的显著区域检测法常常不能准确地探测出显著对象且计算费时。近来,卷积 神经网络模型在图像分析和处理 领域取得了极大成功。为提高图像显著区域检测性能,本文提出了一种基于监督式生成对抗 网络的图像显著性检测方法。它 利用深度卷积神经网络构建监督式生成对抗网络,经生成器网络与鉴别器网络的不断相互对 抗训练,使卷积网络准确学习到 图像显著区域的特征,进而使生成器输出精确的显著对象分布图。同时,本文将网络自身误 差和生成器输出与真值图间的 L1距离相结合,来定义监督式生成对抗网络的损失函数,提升了显著区域检测精度。在MSRA 10K与ECSSD数据库上的实 验结果表明,本文方法 分别获得了94.19%与96.24%的准确率和93.99%与90.13%的召回率,F -Measure值也高达94.15%与94.76%,优于先 前常用的显著性检测模型。

关 键 词:深度学习  显著性检测  生成对抗网络  损失函数
收稿时间:2018/5/7 0:00:00

Image salient object detection based on deep learning
LIU Chun-hui,Zhou Yang,LIU Xiao-qi and TANG Xiang-hong.Image salient object detection based on deep learning[J].Journal of Optoelectronics·laser,2019,30(1):95-103.
Authors:LIU Chun-hui  Zhou Yang  LIU Xiao-qi and TANG Xiang-hong
Affiliation:Faculty of Communication,Hangzhou Dianzi University,Hangzhou 310018,China,Faculty of Communication,Hangzhou Dianzi University,Hangzhou 310018,China,Faculty of Communication,Hangzhou Dianzi University,Hangzhou 310018,China and Faculty of Communication,Hangzhou Dianzi University,Hangzhou 310018,China
Abstract:Salient region detection has been widely applied in object recognition ,image segmentation,and image/video compression. It has an important topic in computer vision.However,the saliency detection me thods based on different visual salient features often cannot accurately detect salient objects,and they are also time-consuming. Recently,the convolutional neural networks (CNN)-based models have achieved remarkable success in image processing and image analysis.In orde r to enhance the detection performance of salient objects,this paper presents an image salient object detection approach based on conditional generative adversarial networks (cGAN).This method uses deep CNN to construct the cGAN.By iteratively fighting each other b etween the generator and the discriminator,the CNN can obtain salient features of images accurately and then urges the generator to produce the high-quality saliency maps.To further improve the saliency detection performance,the loss function of cGAN is defined by cooperating the GAN errors and the L1distance between the generator outputs and ground truth maps.Experimental results on the MSRA10K and ECSSD datasets show that the proposed approach outperforms state-of-the-art saliency detection models significantly ,and has superior performance with 94.19% and 96.24% in precision,93.99% and 90.13% in recall rate,and 94.15% and 94.76% in F-measure .
Keywords:deep learning  saliency detection  generative adversarial networks  loss functio n
本文献已被 万方数据 等数据库收录!
点击此处可从《光电子.激光》浏览原始摘要信息
点击此处可从《光电子.激光》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号