【作者】 李鑫；

【导师】 陈雷霆；

【作者基本信息】 电子科技大学 ， 计算机应用技术， 2018， 博士

【摘要】 人类作为智慧生物,在看到一个场景时,我们的注意力往往会集中在一些重要的对象上,从而自动过滤掉大量不重要的信息。这些具有视觉吸引力的重要对象被称为显著性对象。快速准确地检测出视觉显著的对象可以有效地指导计算资源的分配,排除图像中的干扰因素,并且使得很多复杂的视觉任务变得相对简单。显著性对象在计算机视觉和人工智能领域的各种任务中具有重要的意义,但是准确检测出给定场景的显著性对象却是一个具有很大挑战的任务。首先,显著性对象具有多样性的特征,这种多样性很难用一个通用的数学模型来表达。其次,显著性对象具有对象属性,需要更加语义性的特征来描述。最后,显著性对象检测往往作为其他视觉任务的预处理部分,因而显著性对象检测算法还应该满足高效性。如何让计算机具备与人类相似的视觉处理能力,准确并高效地检测出给定场景中的显著性对象,是当前计算机视觉与人工智能领域的重要课题。本文围绕显著性对象检测算法的若干关键技术,研究了基于图像稠密匹配的显著性对象检测算法,基于全监督深度学习的显著性对象检测方法,以及基于网络监督的显著性对象检测算法,并且讨论了显著性对象检测算法在多种视觉任务中的应用。本文主要创新点包括:1.提出一种基于图像稠密匹配的显著性对象检测算法。不同于传统非监督显著性对象检测算法,该算法从一个新的角度分析和解释图像中对象的显著性。总的来说,该算法基于抽象的视觉显著性特征,通过快速建立与已有标注图像的稠密对应关系,来类比推理出给定图像的显著性对象区域。由于同时考虑了外部样本信息和给定图像的内在信息,使得该方法能够快速准确地检测出给定图像的显著性对象区域。在6个公共数据集上的实验表明,相比已有的传统非监督的显著性对象检测算法,该算法具有更好的效果。2.提出一种基于全监督深度卷积网络的显著性对象检测算法。深度神经网络能够以端到端的方式学习强力的语义特征,从而能极大地提高显著性对象检测的准确性。该算法基于一个新颖的多尺度级联网络模型,以数据驱动的方式学习对象的显著性。不同于已有基于深度学习的显著性对象检测算法,该模型能有效地整合图像的上下文语义信息,并且将后处理部分也集成在了神经网络中,以由粗到精的方式检测显著性对象并不断优化检测结果。在多个公共数据集的实验结果表明,相比现存基于深度学习的显著性对象检测算法,该方法取得了更高的准确率,并且运行效率也优于大多数现存方法。3.深度学习需要运用大量带标注的样本来学习图像的语义特征,因此需要消耗大量的人力、时间以及资金对数据进行标注。为了解决的这个问题,本文提出一种基于网络监督的显著性对象检测方法。该方法利用大量网络上简单易得的、廉价的图像来训练深度神经网络,而无需使用任何的带有显著性对象区域标注的样本进行模型的网络参数训练。实验结果表明,该方法在无需显著性对象标注的情况下,取得了可以媲美基于全监督的深度神经网络模型的效果。4.本文讨论了显著性对象检测算法在两种计算机视觉任务中的应用,包括胸片图像肺部区域分割以及稠密语义匹配。实验结果表明,显著性检测算法可以快速提取视觉显著区域,从而提高计算效率和准确率。更多还原

【Abstract】 When seeing a scenario,we humans,as intelligent beings,often focus our attention on some important objects while unconsciously filter out a lot of unimportant information.These objects that attract our visual attention are called salient objects.Fast and accurate detection of these visually salient objects in images can effectively provide guidance for computer resource allocation,eliminate distraction factors,and make many complex visual tasks easier.Salient objects are of great significance to computer vision and artificial intelligence tasks.However,it is a huge challenge to accurately detect salient objects in a given scenario.First,salient objects have a variety of features that cannot be represented by a simply mathematical model.Second,salient objects have object attributes.Hence,more semantically meaningful features are needed to describe these attributes.Finally,salient object detection often serves as the pre-processing part of other visual tasks.Therefore,salient object detection should be highly efficient.How to equip computers with visual processing attributes similar to humans and how to quickly detect salient objects in a given scenario are an important topic in modern computer vision and artificial intelligence fields.By centering on several key technologies of salient object detection algorithms,this paper researches salient object detection algorithms based on image dense correspondence,analyzes salient object detection algorithms based on fully-supervised deep learning and webly supervised learning,and discusses the application of salient object detection algorithms in various visual tasks.The contributions of this paper are as follows.1.A salient object detection algorithm based on image dense correspondence is proposed.Different from the existing unsupervised algorithms,this algorithm analyzes and explains object saliency in images from a brand new perspective and is able to combine both the intrinsic and extrinsic information.Generally,this algorithm is based on abstract visually salient features and quickly establishes dense correspondence between a given image and the existing annotated images.Using the obtained information,it is able to infer the salient object region of the given image.By considering both the information of external samples and the intrinsic information of the given image,this algorithm can quickly and accurately detect the salient object region in a given image.This algorithm has been tested on sixwidely-used public datasets.The results show that it achieves higher accuracy than the existing unsupervised salient object detection algorithms.2.A salient object detection algorithm based on fully-supervised deep convolutional networks is proposed.The deep neural network is able to learn semantic features in an end-to-end manner.Therefore,it greatly improves the accuracy of salient object detection.This algorithm adopts a novel multi-scale cascaded deep convolutional network and learns the saliency in an end-to-end manner.Different from the existing deep-learning-based saliency object detection algorithms,this algorithm 1)considers multi-scale context information in salient object detection tasks,2)integrates the post processing part into a neural network.This method detects salient objects in a coarse-to-fine manner and constantly adjusts the detection results.The experiments on multiple public datasets show that the proposed algorithm achieves the currently highest accuracy among the existing salient object detection methods.Meanwhile,its operation efficiency is higher than most deep-learning-based object detection algorithms.3.Deep learning requires great numbers of annotated samples to learn the semantic features of images.Therefore,it needs huge amounts of labor,time and capital for data annotation.To solve this problem,this paper proposes a webly supervised salient object detection method.This method utilizes easily accessible images collected from the Internet to train the deep neural network and requires no manual salient object region annotation.Experiments show that this method achieves accuracy comparable to and even higher than fully-supervised deep neural networks.4.This paper discusses the application of object detection algorithms in two computer vision tasks,including region segmentation of sternum X-ray and dense semantic matching.Experiments show that the salient object detection algorithm is able to quickly extract visually silent region,and thus improve computational efficiency and accuracy.更多还原