基于曲线演化的目标分割和识别算法研究
文献类型：学位论文
| 作者 | 范慧杰
|
| 学位类别 | 博士 |
| 答辩日期 | 2013-11-20 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 导师 | 唐延东 |
| 关键词 | 图像分割 目标识别 主动轮廓 曲线演化 水平集函数 灰度不均匀 部分形状描述子 |
| 其他题名 | Object Segmentation and Recognition Based on Curve Evolution |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 图像分割和目标识别是计算机视觉的基础任务，它们作为图像处理的重要内容，是图像分析和理解的前提，被广泛应用于图像检索、医疗诊断、目标分类、目标跟踪等领域。形状作为目标最具有视觉感知意义的重要特征之一，为图像分割和目标识别提供了重要信息，使得基于形状的分割和识别算法成为相关领域十分活跃的研究课题，但同时也存在着诸多挑战。目标形状包含了物体的区域和轮廓特征，本文分别从这两个方面出发，对基于区域的主动轮廓分割算法和基于轮廓的部分形状匹配算法进行了深入研究。针对现有算法存在的问题，本文主要完成了以下研究工作：1、提出了一种矩形约束主动轮廓分割模型，该模型将矩形形状约束作为先验知识直接内嵌于演化水平集函数之中，使得分割过程中，演化曲线始终保持矩形形状。我们将该模型应用于文本图像倾斜角计算。通过对多个文本图像数据库的实验表明，矩形约束主动轮廓模型在很大程度上可以弥补底层图像的目标信息不充分这一缺点，在分割得到文本内容区域的同时给出了目标的定量描述，如文本倾斜角度，内容区域的大小、位置等。2、很多图像中的目标或背景呈现灰度不均匀特性，这会给分割算法带来一定的困难。为了克服灰度不均匀这一问题，我们提出一种全局和局部能量相结合的主动轮廓分割模型，其中全局项收集目标的区域灰度信息，牵引演化曲线向目标位置移动，抑制过分割现象，保证分割算法的稳定性；局部项利用一个Gaussian核函数提取图像的局部灰度信息，控制演化曲线分割目标细节的精细程度。根据目标和背景呈现的灰度不均匀特征的不同，在实际应用选取不同的局部项，使得分割结果更加精确。3、曲率作为曲线的重要特征之一，其计算的精度直接影响对目标形状的描述。我们以国家自然科学基金“基于水平集的PLIF火焰前锋与特征提取算法研究”为研究背景，提出基于水平集的火焰前锋识别和曲率计算方法。在保证火焰前锋分割结果符合物理真实的前提下，对火焰前锋形线上点的曲率精确计算，通过对大量的PLIF火焰图像进行实验，得到的实验结论为进一步完善现有的燃烧模型提供了数据支持。4、作为描述目标的重要特征之一的轮廓，在目标被遮挡或自遮挡时会出现断裂、丢失现象，类内对象轮廓存在差异性和类间对象轮廓存在相似性等都会给基于局部轮廓针对这一问题，我们构建了一种新的部分形状描述子并提出了基于单模板的部分形状识别算法。该方法适用于无需任何标记的自然场景目标识别和分类。我们构建的部分形状描述子在很大程度上克服了部分形状匹配算法面临的两个重要问题：1）特征选取不一致性问题，即提取的两个匹配片段的起点不一致或者范围不一致；2）同时解决匹配曲线对的平移、旋转、尺度问题。我们在通用图像数据库ETHZ上验证了算法对目标刚性形变的适用性，而且对目标部分缺失、非刚性形变等情况亦有较好的识别结果。 |
| 索取号 | TP391.41/F23/2013 |
| 英文摘要 | Image segmentation and object recognition are fundamental tasks in computer vision field, as important contents of image processing and premise of image analysis and image understanding, they are widely used in many fields such as image retrieval, medical diagnosis, target classification and target tracking. Shape is considered as one of the most important features with Visual perceptual significance, it also provides important information for image segmentation and object recognition which makes segmentation and recognition algorithms based on shape become very active topics in relative research fields. But meantime, it always presents a lot of challenges. Shape often contains region and contour features in an object, in these two aspects we make further study on region-based active contour segmentation model and contour-based partial shape matching. In order to resolve the problems in current algorithms, we mainly completed the following work:1、We propose a segmentation model of rectangular constrained active contour, in the model, the rectangular shape constraint is regard as a priori knowledge and embedded in evolution level set function directly, which makes evolution curve always keep the rectangular shape. We use this model to detect the inclination angle of text images, experiment results on many text image database indicate that rectangular constrained active contour model can make up for the deficiency of bottom image lack of information. It not only segments and obtains the region of text content, but also presents the quantitative description of given target, for example, inclination angle of text, scale and position of the content area and so on.2、There are intensity inhomogeneous properties within many targets or background in images, which might cause a great difficult to segmentation, to overcome the intensity inhomogeneity, we present a new active contour segmentation model which combines the global and local energy, the global term is used to collect region intensity information and drag the evolution curve moving to target position and restrain over-segment phenomenon to maintain the stability of segmentation, the local term utilizes a Gaussian kernel function to extract the local intensity information of image and controls the fine degree of evolution curve segmenting the detail of target. According to the difference of intensity inhomogeneity between target and background, we choose different local term to leading to more precise segmentation result.3、Curvature is one of important features of curve, its calculation accuracy will influence the shape description directly. We propose a flame front recognition and curvature calculation methods based on the level set function, on the premise of segmentation result of flame front according with physical reality, we calculate the curvature of the flame front contour points precisely and test our methods on plenty of PLIF flame images, the experimental conclusions are helpful to improve and perfect the current combustion model.4、Contour as one of important features of object description can fracture or be missing when object occluded or self-occultation, both the difference of intra-class contours and the similarity of inter-class contours can cause a great challenge on shape matching based on local contour. We proposed a single model based partial shape recognition method. The method applied to object recognition and classification in natural scene without any label. The partial shape descriptor constructed in the model can overcome two major problems that partial shape matching methods facing: 1) feature selection inconsistency, which is the start points and scope of extracted fragment pair are inconsistent; 2) resolve the translation, rotation, scale problems of matching curve pair simultaneously. We verified the applicability of proposed method on rigid deformation in ETHZ database, and it also has good result when part of object missing or non-rigid deformation. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 112页 |
| 分类号 | TP391.41 |
| 源URL | [http://ir.sia.ac.cn/handle/173321/14835]  |
| 专题 | 沈阳自动化研究所_机器人学研究室 |
| 推荐引用方式 GB/T 7714 | 范慧杰. 基于曲线演化的目标分割和识别算法研究[D]. 中国科学院沈阳自动化研究所. 2013. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
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。