复杂场景下的显著性检测与应用

发布时间：2018-10-21 08:24

【摘要】：随着社交媒体、电子商务等的快速发展,图像在信息交流中的作用越来越大,人们对图像信息处理的需求越来越大。人们要求计算机能够更加智能更加自主地实现对图像信息的处理。图像作为数据的一种存储形式,其数据量大,而真正对任务有用的信息只存在于图像的部分区域。人眼可以快速地自发地找到场景中令人感兴趣的目标或者区域,这样的能力正可以协助计算机快速地寻找到能够有效表示图像的目标或者区域。图像显著性检测正是模拟人眼的这种能力而衍生出来的一个科研领域。图像显著性检测借鉴了生物学领域分析人眼的一些理论成果,使用计算的方法从图像中检测令人感兴趣的目标或者区域。经过十数年的发展,图像显著性检测已经获得了巨大的发展,很多图像显著性检测技术在一些数据集上表现出了很高的性能,但是当前的图像显著性检测技术依然无法应对复杂场景以及多显著目标的情况。本文针对这些问题提出了两种方法:(1)基于目标先验的显著性检测,本文先使用目标检测的技术提取多类目标的目标窗口,基于每个目标窗口的前景先验进行显著性检测,然后对所有显著图进行评价,基于评价将基于多类目标先验的显著图进行融合。(2)基于视点预测先验的显著性检测,本文先使用视点预测的技术预测人眼关注区域,并分割前景先验区域,然后基于前景先验区域进行显著性检测。本文在最后对显著性检测技术在目标识别领域的应用进行了研究。研究发现当前的显著性应用或是混合了前景和背景的信息,或是忽视了背景的信息。针对这一问题,本文提出了一个基于显著性的多层模糊表示的目标识别的方法,将背景的信息和前景的信息使用多核学习的方法进行融合,提升了目标识别的精度。
[Abstract]:With the rapid development of social media and electronic commerce, images play more and more important role in information exchange, and people need more and more image information processing. Computers are required to process image information more intelligently and autonomously. As a storage form of data, image has a large amount of data, but the information that is useful to the task only exists in some areas of the image. The human eye can quickly and spontaneously find the interesting target or region in the scene. This ability can help the computer to quickly find the object or region that can effectively represent the image. Image salience detection is a scientific research field that simulates this ability of human eyes. Image salience detection draws lessons from some theoretical achievements of analyzing human eyes in the field of biology and uses computational methods to detect objects or regions of interest from images. After more than ten years of development, image salience detection has made great progress, and many image salience detection techniques have shown high performance in some data sets. However, the current image salience detection technology still can not deal with complex scenes and multiple salient targets. In order to solve these problems, two methods are proposed in this paper: (1) based on the priori salience detection of the target, this paper first uses the technology of target detection to extract the multi-class target window, and carries out the salience detection based on the foreground priori of each target window. Then all the salient maps are evaluated. Based on the evaluation, the salience map based on multi-class objective priori is fused. (2) based on the significance detection of the viewpoint prediction priori, this paper first uses the viewpoint prediction technology to predict the region of human eye concern. The foreground priori region is segmented and the significance is detected based on the foreground priori region. At the end of this paper, the application of salience detection in target recognition is studied. The study found that current salient applications either mix foreground and background information or ignore background information. In order to solve this problem, a method of target recognition based on salient multi-layer fuzzy representation is proposed in this paper. The background information and foreground information are fused by multi-core learning method to improve the accuracy of target recognition.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】