毕业生知识库

近年来，互联网技术的飞速发展和手持移动设备的大众化，使得生产和创作图像变得越来越简单。社交网络的不断发展，直观而生动的图片逐渐取代了繁琐而微妙的文字，成为传词达意的主要媒介。与此同时，图像数据量呈爆炸性增长，人们对智能图像内容分析、智能图像编辑的需求与日俱增。然而，大部分传统的图像理解和处理技术仍然停留在像素、超像素等低层阶段，无法从语义层面响应人类的真实意图。如何跨越低层图像表示和高层图像语义之间的“语义鸿沟”，实现图像内容的高层语义理解和智能编辑，已经成为图像处理、计算机视觉、计算机图形学等领域的重要研究课题。 

本文围绕图像内容分析、理解和编辑这三个重要问题，以图像中大量存在的相似模式为研究对象，研究了图像中相似物体的检测、图像中相似物体的重要度测量与估计和基于相似模式分析的图像缩放三个问题。具体地，本文主要的研究工作如下： 

1.提出一种基于简单交互的多特征融合的相似物体检测方法。该方法联合多种视觉特征，在多特征独立匹配的结果之上，采用一种新的优化融合策略对特征的权重进行动态调整，可以更加可靠地利用不同特征的优点，减少匹配错误的引入。经过实验证明，本方法在适用场景和鲁棒性方面优于已有方法。 

2.提出一种图像中相似物体的重要度测量和估计方法。在重要度测量方面，我们建立了物体重要度和物体排序序列之间的概率统计模型，基于最大似然方法从同一幅图像的多个物体排序序列中对物体的重要度进行测量，并建立了一个包含808张图像的相似物体图像数据库来验证我们方法的可行性。在重要度估计方面，我们提出多种简单有效的视觉特征，通过排序学习和随机森林方法，对相似物体的重要度进行有效地估计。最后，我们通过多种应用证明了本方法的有效性和实用性，与显著性方法的对比实验证明，本方法在物体级别的重要度估计上显著好于基于低层特征的显著性方法。 

3.提出一种基于相似模式分析的图像缩放方法。该方法对图像中的相似物体进行检测和重要度估计，并提出新的“物体焊接”缩放算子。在图像缩放的过程中，通过提出的能量评估函数，在经典的图像缩放算子和物体焊接算子间进行选择，可以自动地移除场景中的次要相似物体，在避免图像内容变形的同时保留图像的语义结构。实验结果表明，本方法在继承已有图像缩放方法优点的同时，在缩放包含相似物体的场景时可以得到更加美观合理的缩放结果。

In recent years, with the rapid development of Internet and the popularity of mobile device, it becomes more and more simple to produce and create photos. In a variety of social networks, intuitive and vivid images gradually replace the complicated and delicate words, become the main media to convey feeling and ideas. Meanwhile, with explosive growth of the image data, the demand of intelligent image content analysis and editing is steadily on the increase. However, most of the traditional image understanding and processing technology at pixel or superpixel level, unable to response the real intention of humans from the semantic level. How to step over the "semantic gap" between low-level image representation and high-level image semantics, understanding and editing image through an intelligent way, has become an important research topic of image processing, computer vision, computer graphics and other fields. 

 

In this thesis, we focus on image content analysis, understanding and editing. More specifically, we mainly study similar objects detection and segmentation, measuring and predicting visual importance of similar objects, and similar pattern based image retargeting. The major contributions of this thesis are: 

1. We propose a novel multi-feature based similar objects detection method. This method where user scribbles are used to guided detection and analysis of similar objects, uses an optimization method to integrate of multiple features more effectively. The experiments show that our algorithm is more robust for a variety of scenes than the previous work. 

2. We propose an approach to measure and predict visual importance of similar objects. We have established a probabilistic model between the object importance and the object permutations, and use the maximum likelihood method to measure similar objects importance from a plurality of object sequence labeled by users. Furthermore, we propose a variety of simple and effective visual features, and predict importance of similar objects by learning to rank method. A dataset consisting of 808 images with ground-truth annotations is also introduced to evaluate the result. The experiments show that our method outperforms state-of-the-art saliency methods. 

3. We propose a similar pattern analysis based image retargeting method. The method first detects the similar objects and predicts importance. With an energy evaluation function, this method removes least importance object automatically through an object carving operator during the resizing process. We demonstrated that our method has distinct advantages to resize the scene containing similar objects.