基于双层多尺度神经网络的显著性对象检测算法  

Salient Object Detection Algorithm Based on Dual-Layer Multi-Scale Neural Network

作　　者：李鑫[1] 陈雷霆[1,2,3] 蔡洪斌 李建平[1] 杨帆[1] LI Xin;CHEN Lei-ting;CAI Hong-bing;LI Jian-ping;YANG Fan(School of Computer Science ＆ Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China;Digital Media Technology Key Laboratory of Sichuan Province,University of Electronic Science and Technology of China,Chengdu 611731,China;Institute of Electronic ＆ Information Engineering in Guangdong,University of Electronic Science and Technology of China,Dongguan 523000,China)

机构地区：[1]电子科技大学计算机科学与工程学院,四川成都611731 [2]电子科技大学数字媒体技术四川省重点实验室,四川成都611731 [3]电子科技大学广东电子信息工程研究院,广东东莞523000

出　　处：《微电子学与计算机》2018年第11期1-7,共7页Microelectronics & Computer

基　　金：国家自然科学基金(61370073);国家863计划项目(2015AA016010);广东省科技计划项目(2016A040403004)

摘　　要：为了提高显著性对象检测的准确率,本文提出一种基于双层多尺度神经网络的深度模型.不同于现有的深度神经网络模型.首先,该模型以由精到粗的方式进行深度特征学习,并且定位显著性对象的初始位置;然后,以由粗到精的方式整合多尺度上下文语义信息,从而精确检测整个显著性对象区域,输出相应的显著性图;最后,为了进一步提高检测结果的准确率,利用全连接条件随机场对输出的显著性图进行优化,得到最终的显著性对象检测结果.在多个显著性对象检测公共数据集的验证结果表明,本文算法在运行效率和准确率上均优于当前传统显著性对象检测算法以及现有的基于深度学习的显著性对象检测算法.To further improve the accuracy of salient object detection,a novel Dual-Layer Multi-Scale Neural Network（DLMSNN）was proposed in this work.Different from existing deep model,the proposed deep model first learned deep features in a fine-to-coarse manner,and roughly located the salient object regions.Then,the model integrated multi-scale contextual information in a coarse-to-fine manner to precisely detect the entire salient object regions,and generated accurate saliency map of the input image.Finally,to further improve the performance,the dense conditional random field algorithm was used to refine the saliency map and produce the final result.The experimental results on several public benchmarks showed that the proposed algorithm outperformed traditional salient object detection methods and existing deep learning-based algorithms.

关 键 词：显著性对象检测 深度学习 深度卷积网络 条件随机场 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]