基于深度学习的复杂图像显著物体检测方法研究

项目名称： 基于深度学习的复杂图像显著物体检测方法研究

项目编号： No.61473231

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 韩军伟

作者单位： 西北工业大学

项目金额： 82万元

中文摘要： 显著物体检测技术能够自动提取场景中的重要内容，是图像理解中的一个基础问题，可广泛的应用于图像/视频编码与传输、图像搜索与识别、视频监控、遥感图像监测等领域。针对已有方法对复杂图像处理效果不佳的现状，本项目提出利用深度学习算法从数据中挖掘能够刻画复杂图像内在模式的特征，以此为基础构建显著物体检测模型的新思路。结合申请人前期工作（研究成果发表在国际期刊IEEE TCSVT,IEEE TIP等），重点研究以下四个主要内容：1）构造非监督深度学习算法对复杂背景建模来准确提取前景与背景间对比度信息；2）基于有监督深度学习算法学习物体属性特征，与对比度信息融合实现显著物体检测；3）构建用于显著物体检测的大规模复杂图像测试库；4）在图像压缩、互联网图片分类搜索、移动终端图片显示应用中对方法进行验证和应用。本项目采用国际合作研究方式，力争在算法理论和应用方面取得创新成果，并推动图像理解等相关领域的发展。

中文关键词： 视觉显著物体检测；深度学习；对比度特征；物体属性

英文摘要： Salient object detection enables to automatically infer the important content in a visual scene, which is a fundamental problem in image understanding area. It can be widely used in the applications such as image/video coding, image search and recognition, video surveillance, and remote sensing image surveillance. Based on the fact that current methods can not handle complex images well, this project proposes a novel idea which builds saliency detection model based on features learned from image data. The learned feature is able to capture hidden patterns in complex images. Based on our precious works published on IEEE TCSVT and IEEE TIP, this project focuses on four major techniques: 1) Develop unsupervised deep learning algorithm to model the background and thus accurately obtain the contrast between background and foreground; 2) Develop supervised deep learning algorithms to learn object attribute features and combine with contrast to detect salient objects; 3) Construct a large-scale dataset containing complex images for algorithm testing and comparison; 4) Validate and apply the proposed method in applications of image coding, image search engine, and mobile image visualization.This project will be conducted in a manner of international cooperation. It aims for innovative achievements in both theory and application and will also be very attractive for other relevant domains of image understanding.

英文关键词： Visual salient object detection;Deep learning;Contrast feature;Objectness attribute