CASIA OpenIR  > 毕业生  > 博士学位论文
鲁棒的目标检测与识别方法研究
其他题名Robust Approach for Object Detection and Recognition
洪义平
学位类型工学博士
导师易建强
2005-04-01
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业控制理论与控制工程
关键词目标检测 字符识别 分割 Object Detection Character Recognition Segmentation
摘要随着计算机技术的发展,图像目标的检测与识别己在很多领域得到了广泛的应用。目标检测与识别存在的关键问题是分割,即如何从复杂的背景中提取出目标,且具有较好的鲁棒性。本文主要研究了复杂背景下的鲁棒目标分割方法、目标检测与识别机制、以及用于移动机器人导航的实际目标检测与识别。本文的主要研究成果可归纳如下: 1) 研究并提出了一种鲁棒的自然图像分割方法。在自然图像分割中,同时考虑纹理和轮廓颜色信息的分割方法大多是基于递归处理的,计算时间长,无法适应具有实时性要求的场合。为了适应视觉处理和实时性的要求,在 Mean Shift 方法的基础上提出了一种改进的分割方法。实验结果表明,对不同光照条件下和具有纹理的自然图像,都能获得稳定的分割结果,鲁棒性能较好。而且,可根据上层视觉的需要调整参数,来获得所需的分割结果。 2) 在鲁棒自然图像分割方法的基础上,提出了一种由粗到精的目标检测与识别机制。对任意采集的一幅图像,首先用鲁棒自然图像分割方法进行粗分割。当图像中主要包含所要识别的目标时,再采用精处理的方法。在分割之后,根据已建立的目标模型进行认证识别。 3) 设计了一种基于鲁棒分割方法和目标颜色建模的门和道路检测方法。首先用鲁棒自然图像分割方法将图像分割成一些区域,然后对每个区域计算颜色平均值,用离线建立的目标颜色模板来认证目标候选区域。 4) 设计了一种自然场景下的字符识别系统(门牌号自动识别系统)。这一识别系统具有以下特点:较好的抗噪声能力、较好的实时性和较好的可推广性。本文的鲁棒自然图像分割方法具有较好的分割效果,且计算时间短,可用于具有实时性要求的场合。门和道路的检测以及门牌号识别系统可推广到机器人其它目标识别应用中。
其他摘要Detection and recognition of objects in images has been applied in many areas. The key point in object detection or recognition is the image segmentation. How to robustly extract the interesting objects from complex background will greatly determine the final detection or recognition result. Aiming at developing a vision system for the robot based on natural object detection or recognition, the approach of robust segmentation, the scheme of object detection and recognition, and some real examples of object detection and recognition in mobile robot navigation are researched. The main contributions of the works reported in this thesis are as follows: 1. A robust approach for natural image segmentation is proposed. Most of the existing methods for natural image segmentation simultaneously considering the texture and contour features are based on the recursive process that is much time-consuming. Therefore, they are not suitable to the cases requiring real-time processing. In order to apply the segmentation methods for natural images in the real-time applications, an approach derived from Mean Shift is developed. The experimental results show that the different images under various illuminations and backgrounds can be stably partitioned into some meaningful regions. In addition, the parameters used in segmentation are adjusted by the real vision task. 2. Based on the proposed robust approach for natural image segmentation, a new scheme of object detection and recognition is presented. The new scheme includes following important steps: roughly segmenting the images with big parameters in the proposed natural image segmentation, extracting the interesting regions with meticulous image processing, and verifying or recognizing the object candidates with object models. 3. A detection approach for the door and the floor used in robot navigation is designed. The detection procedure is based on the robust natural image segmentation and the robust statistical modeling of objects. 4. A character recognition system under natural scenes used in robot navigation is designed. Compared with other character recognition systems for robot navigation, this system is much robust and resistant to the noise. The proposed approach for natural image segmentation has not only a robust result, but also a fast computing speed which makes it can be used in robot navigation. The detection or recognition methods of the door, the floor and the characters in natural scene can also be used in the other object detection or recognition.
馆藏号XWLW934
其他标识符200218014603160
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/5842
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
洪义平. 鲁棒的目标检测与识别方法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2005.
条目包含的文件
条目无相关文件。
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[洪义平]的文章
百度学术
百度学术中相似的文章
[洪义平]的文章
必应学术
必应学术中相似的文章
[洪义平]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。