视觉显著性物体检测方法及应用研究

发布时间：2018-05-21 05:58

本文选题：显著性物体检测 + 显著性偏置　；参考：《中国科学技术大学》2016年博士论文

【摘要】：近年来,图像的数量随着移动互联网的发展呈现爆发式的增长。为了从海量图像数据中寻找自己所需的信息,人们迫切需要快速、准确的图像处理技术。人类视觉系统为了解决大脑处理能力有限的问题,在接收场景信息时会选择重要的视觉信息进行优先处理。这种选择性注意机制使人类能够快速适应外界的变化。受此机制的启发,研究人员提出了视觉显著性检测方法来模拟人类的视觉注意机制。显著性检测算法能够定位图像中吸引人注意力的区域,非常适合用于排除图像中无关内容的干扰,从而大大加快传统图像处理的速度。显著性检测目前已经成为计算机视觉的热门研究方向,并广泛应用在多个计算机视觉领域中,如目标分割、物体识别和目标跟踪等。人类视觉系统在观察环境时可以分为快速的、与任务无关的自底向上的方式和慢速的、由任务驱动的自顶向下的方式。由于自底向上的方式不需要高层知识的引导,因此大多数研究关注自底向上的显著性检测。本文主要针对自底向上的显著性物体检测展开研究,通过分析已有的显著性算法的缺点,并结合显著性的生物学原理,提出了两种新颖的显著性物体检测算法,并成功应用在目标分割等计算机视觉应用中。本文的主要研究工作和贡献可概括如下：1)基于人类视觉系统的显著性物体选择机制,提出了基于显著性偏置的显著性物体检测算法,将区域显著性计算和物体性计算明确区分开。该方法首先计算每个区域属于物体的概率(即物体性)来定位图像中所有可能的物体区域,然后基于对比度计算每个区域的显著性,最后通过非线性融合的方式实现区域显著性对物体性的偏置,得到显著性物体区域。为了解决同类区域显著值不一致的问题,提出了基于显著性扩散的优化方法,从初始显著图中选择种子点,并利用区域特征学习区域间相似度,然后根据其他区域与种子点的相似性关系优化每个区域的显著值,得到更加一致的显著性检测结果。实验结果验证了所提算法的有效性。2)基于区域显著性产生的生物学原理,提出了背景驱动的显著性检测算法。通过分析已有的基于局部或全局对比度的显著性检测方法的局限性,发现背景在对比度计算中的重要作用,并从背景图中分割出背景区域作为对比度计算参考区域。背景图可以利用任何背景先验得到,我们特别提出了基于卷积神经网络的背景学习模型来预测每个区域属于背景的概率。计算区域对比度时采用颜色和纹理特征,并且根据特征的分布情况动态确定两者的权重。为了提高显著性物体的完整性,提出了基于增强图模型的优化方法,在传统的k-正则图中嵌入背景先验,并添加特征空间中的非局部连接,然后利用节点间相似度在图上传播并优化显著值。在多个典型数据集上的实验结果证明了所提算法的有效性。3)为了验证显著性物体检测的应用价值,将所提出的检测算法应用于目标分割和物体分类中。在目标分割中,探讨了显著图的自适应分割、基于GrabCut的分割和基于用户交互的分割方法,对比实验结果表明显著性检测可以促进目标分割的效果。在物体分类中,为了排除背景区域特征的干扰,在显著图分割出的前景区域中提取特征并进行分类。通过比较不同显著性检测算法对分类性能的影响,表明显著性检测可以增强物体分类的性能,并且检测效果越好,分类性能也越强。总结起来,本文针对显著性物体检测方法进行了深入研究,以显著性产生的相关生物学原理为指导,提出了两种显著性物体检测方法,即基于显著性偏置和扩散的方法和背景驱动的方法,在提高检测物体的显著性和完整性方面取得了领先的效果。在显著性检测应用中,探讨了显著性物体检测在目标分割和物体分类中的应用,展示了其应用价值。
[Abstract]:In recent years, the number of images has been increasing with the development of mobile Internet. In order to find out the information needed from the massive image data, people urgently need fast and accurate image processing technology. In order to solve the problem of limited brain processing ability, human visual system will choose important information in receiving scene information. Based on this mechanism, the researchers have proposed a visual significance detection method to simulate human visual attention mechanism. In addition to the interference of unrelated content in the image, the speed of traditional image processing is greatly accelerated. Significant detection has become a hot research direction in computer vision and is widely used in the field of multiple computer vision, such as target segmentation, object recognition and target tracking. Fast, bottom-up and slow, task driven, top-down, task driven, top-down way. Most studies focus on bottom-up saliency detection because of the bottom-up approach. Therefore, this paper focuses on bottom-up significant object detection. With the shortcomings of some significant algorithms, and combining with the biological principles of significance, two novel detection algorithms for significant objects are proposed and applied to computer vision applications such as target segmentation. The main research work and contributions of this paper are summarized as follows: 1) the significant object selection mechanism based on human visual system is proposed. A significant object detection algorithm based on significant bias, which separates the regional saliency calculation from the object calculation. This method first calculates the probability of each area belonging to the object (i.e. the object) to locate all the possible object regions in the image, then calculates the saliency of each region based on the contrast, and finally passes the nonlinearity. In order to solve the problem of congenialization, an optimization method based on significant diffusion is proposed to select the seed points from the initial significant map, and to learn the similarity between the regions and then according to the other regions and species. The similarity relation of the subpoints optimizes the significant values of each region and obtains a more consistent significance detection result. The experimental results verify the validity of the proposed algorithm.2) based on the biological principle generated by the regional significance, the background driven significance detection algorithm is proposed. The limitations of the method of sex detection, find the important role of the background in the contrast calculation, and divide the background area into the contrast calculation reference area from the background map. The background map can be obtained by any background prior. We specially propose a backview learning model based on the convolution neural network to predict the background of each region. In order to improve the integrity of the significant objects, an optimization method based on the enhancement graph model is proposed to embed the background pre test in the traditional k- regular graph and add the non local connection in the feature space. The experimental results on multiple typical datasets prove the validity of the proposed algorithm.3). In order to verify the application value of the significant object detection, the proposed detection algorithm is applied to the target segmentation and object classification. The segmentation, GrabCut based segmentation and the user interaction based segmentation method, the experimental results show that significant detection can promote the effect of target segmentation. In object classification, in order to eliminate the interference of the background region features, the features are extracted and classified in the foreground region divided by the significant graph. By comparing the different saliency of the object classification. The effect of detection algorithm on classification performance shows that significant detection can enhance the performance of object classification, and the better the detection effect, the better the classification performance. In this paper, this paper has carried out a thorough study on the detection methods of significant objects, guided by the principle of significant related biological science, and proposed two kinds of significant physical examination. The method, which is based on the method of significant bias and diffusion and the method of background driven, has achieved the leading effect in improving the significance and integrity of the object detection. In the application of significant detection, the application of significant object detection in target segmentation and object classification is discussed, and its application value is demonstrated.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】