基于图模型的图像显著性目标检测理论与方法研究

Salient object detection of image has got rapid development over the last few years. Due to the high accuracy, graph based Salient object detection methods have attracted more and more attention. Since there are many images with complex structure and color distribution in real world, we face many research problems and challenges in salient object detection and its related applications. In this project, we focus on graph based image salient object detection and intend to study rational multiple feature fusion methods; Establish effective graph learning based on image region self-similarity; Propose initial background seeds selecting strategy and optimize semi-supervised learning algorithm. On the bases of those researches, we aim to solve the key theory and technology for graph and semi-supervised learning based salient object detection. It is expected that this project will provide theoretical basis and technical support for bearing information business in our country.

图像的显著性目标检测技术近年来得到了迅速发展，其中基于图模型的显著性目标检测方法以其高的检测准确率得到了越来越多研究者的关注。由于现实生活中图像结构和颜色分布的复杂性，在图像显著性目标检测和其相关应用上面临着新的问题和挑战。本项目围绕基于图模型的图像显著性目标检测方法，拟研究合理的的图像多特征融合方法；建立有效的基于图像区域自相似性的图学习模型；研究更加精确的图像背景种子选取方法和优化的半监督学习算法。通过以上研究，有效解决基于图模型的图像显著性目标检测中图模型的优化问题和适用于图像显著性目标检测中半监督学习的主要理论问题，为我国多媒体信息技术产业的发展奠定理论基础和技术支撑。

图像的显著性目标检测技术近年来得到了迅速发展，其中基于图模型的显著性目标检测方法以其高的检测准确率得到了越来越多研究者的关注。由于现实生活中图像结构和颜色分布的复杂性，在图像显著性目标检测和其相关应用上面临着新的问题和挑战。本项目围绕基于图模型的图像显著性目标检测方法，提出了合理的的图像多特征融合方法；建立了有效的基于图像区域自相似性的图学习模型；研究了更加精确的图像背景种子选取方法和优化的半监督学习算法；提出了一套有效的深度神经网络特征融合方法；通过该项目的研究，有效地解决基于图模型的图像显著性目标检测中图模型的优化问题和适用于图像显著性目标检测中半监督学习的主要理论问题，为我国多媒体信息技术产业的发展奠定理论基础和技术支撑。