CASIA OpenIR  > 智能感知与计算研究中心
基于深度强化学习裁剪物体背景的弱监督物体定位
郑武
Subtype硕士
Thesis Advisor张兆翔
2019-06
Degree Grantor中国科学院大学
Place of Conferral北京
Degree Discipline计算机技术
Keyword弱监督物体定位,深度强化学习,背景裁剪
Abstract

弱监督物体定位是视觉定位中的一类特殊问题,这类任务往往只提供图片的类别标签作为监督信息,区别于强监督定位,后者同时提供图片标签和物体边界框作为监督信息。因此,在监督信息更少的情况下,弱监督物体定位的难度远高于强监督物体定位。然而,更少的监督信息意味着对人工标注数据集的依赖更弱,因而弱监督物体定位可以节省数据标注所花费的大量人力与金钱,并提升弱监督定位算法和模型对于更多数据集的适用性。因此,弱监督物体定位成为近年来的研究热点之一,吸引了学界广泛的注意力。为此,本文提出了一种基于深度强化学习、通过裁剪物体背景的弱监督物体定位方法,并提出了一种新颖的修正方法来进一步提升模型的定位性能,在常用的物体定位数据集上取得了良好的效果。本文的主要创新点包括:
    提出了一种通过裁剪物体背景实现物体定位的方法,将物体背景裁剪过程建模成马尔科夫决策序列,并采用深度强化学习算法寻求最优解,该方法更加符合人类视觉机制,且定位效率较高。
    提出了一种基于卷积特征图重组的弱监督定位修正方法,该方法通过重组图片所提取的特征图,利用特征图中的空间位置信息,校正深度强化学习框架的定位结果,进一步提升了整体框架的定位性能。
    提出了一种可移植性较强的定位框架,易于将其嵌入到已有的卷积分类网络中,同时实现分类和定位任务。

最后本文总结了以上提出的方法中的关键问题和应对策略,以及未来亟待探索的研究方向。

Other Abstract

Weakly-supervised object localization is a kind of special problem in visual localization, because it is usually only provided with image-level labels in the task, which is different from fully-supervised object localization, the latter task could have both image-level labels and bounding box annotations as supervision information. Thus, with less supervision, weakly-supervised object localization is more difficult than the fully-supervised one. However, less supervision also means less dependency on human-labour to annotate the datasets, which can save large amounts of human-labour and money, as well as improve the adaptability of weakly-supervised object localization algorithms and frameworks on more datasets. Recently, weakly-supervised object localization became a hot research topic and attracted much attention. Therefore, we propose a deep reinforcement learning approach to achieve weakly-supervised object localization by cutting background, and also design a novel refinement method to further improve the localization performance, which achieves good results on popular datasets. The main contributions of this work can be summarized as follows:
    We propose a localization method by cutting object background and model the cutting process as a Markov decision process, which can be solved by the designed deep reinforcement learning algorithm. The proposed method complies more with human visual mechanism and has higher efficiency.
    We propose a localization refinement method by re-organizing the convolution feature maps. This method integrates the location information in the feature maps and takes it as complementarity for the deep reinforcement learning framework, with localization performance improved significantly.
    We proposed a localization framework that can easily be embedded into existing convolutional neural networks to realize localization and classification simultaneously.

In the end, we summarized the key issues and coping strategies in the proposed methods as well as the research directions that need to be explored in the future.

Pages77
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23861
Collection智能感知与计算研究中心
Recommended Citation
GB/T 7714
郑武. 基于深度强化学习裁剪物体背景的弱监督物体定位[D]. 北京. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(3920KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[郑武]'s Articles
Baidu academic
Similar articles in Baidu academic
[郑武]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[郑武]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.