图像一致性特征与目标检测方法研究
文献类型：学位论文
| 作者 | 蔺蘭 |
| 学位类别 | 博士 |
| 答辩日期 | 2014-11-28 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 中国科学院沈阳自动化研究所 |
| 导师 | 唐延东 |
| 关键词 | 目标检测 一致性 区域一致性 一致性算子 一致性特征融合 |
| 其他题名 | Conformance Feature and Object Detection in Images |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 目标检测作为图像处理的重要内容是计算机视觉的基础任务，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。目标检测的对象和检测方法多种多样，本文以可见光图像为研究对象，针对现存算法复杂性与有效性之间的矛盾，从一致性“趋向相同”的原理出发，依据由一维至三维数据的思路，完成了以下研究工作： 1）介绍了目标检测的研究背景及意义、对象及主要方法，针对方法的复杂性引入一致性的概念，分析了一致性的词义及其在计算机信息科学领域内的含义，叙述了一致性在异常目标检测方面的概况，总结了可用于反映图像一致性的度量。 2）提出了基于区域一致性特征的文本布局检测方法，构建一致性函数来反映局部区域的一致性程度，再从全局角度通过衡量文本区域与背景区域的一致性差异实现文本布局的检测分析。基于一致性特征，该方法具有旋转不变性、语义无关性、抗噪性、背景过滤及多尺度的特点和检测图文混排、手稿、变形文字等多种用途。在MARG、Tobacco和MTDD公用数据集上的实验结果表明了方法的有效性，与多种方法的比较结果显示了方法在适用性、有效性、简便性方面的优势。 3）提出了基于一致性算子的彩色图像显著目标检测方法，构建了基于Lab空间的一致性算子以衡量局部数据元素的一致性程度，通过检测目标与背景一致性的差异得到最终的检测结果。该方法无需进行复杂的特征提取过程，也未采用更多的先验和复杂的分析方法，大大简化了算法过程。在MSRA-1000显著性目标公用数据集上以Ground Truth为标准对算法进行了检验，结果表明了方法的有效性。 4）提出了一种基于重复性特征的绝缘子缺陷检测方法，构建了反映绝缘子图像中结构元素的重复性特征向量，通过该向量可以判断绝缘子是否为缺陷绝缘子，通过检测特征量的野值以检测出绝缘子的缺陷位置。在绝缘子图像数据集和包含绝缘子的无人机视频上的检测结果表明，基于重复特征的检测方法能够快速、准确地检测出有缺陷的绝缘子目标，且可以处理单绝缘子串上的多个缺陷和多绝缘子串的缺陷检测任务。重复性特征是一致性在一维向量上的体现，基于重复性特征的绝缘子缺陷检测方法是一致性思想的具体运用。 5）通过一致性特征与其它特征的融合应用，针对不同的研究问题提出了新的解决方案。基于一致性光照估计和光照系统几何建模的人脸图像光照补偿算法，采用一致性思想估计单幅图像的光照条件，结合光照系统、生物测定学相关知识提出了新的人脸光照模型，用于人脸图像的光照补偿、在大角度光照条件的情况下得到十分理想的识别结果；基于深度一致性的手势识别算法、通过深度的一致性差异、皮肤色彩特征和手势建模，得到了具有方向不变性的手势识别结果；基于均匀一致性的轮胎痕迹种类识别算法，利用轮胎痕迹内部、边界及背景的均匀性特点检测特征点，通过匹配特征点得到轮胎痕迹的候选种属集。以上算法的实验结果均表明了一致性特征融合的可行性及有效性。 |
| 索取号 | TP391.41/L64/2014 |
| 英文摘要 | As an important part of image processing, object detection is the basis task of computer vision and the premise for image analysis and understanding. Object detection is widely used in image retrieval, medical diagnostics, target classification, target tracking and other fields. In this dissertation, we concentrate on the contradiction between complexity and effectiveness of existing algorithms dealing with visible images. Deriving from the meaning of conformance, we completed following works: 1) We introduce the background, significance, research object and major methods of object detection; introduce the concept of conformance, analysis of conformance in the fields of Computer and Information Science; overview conformance in terms of object detection and summarize measurements that can be used to reflect the image conformance. 2) We propose a text layout detection method based on region conformance feature. To reflect the degree of conformance in the local area, we build the conformance function, and then realize the detection and analysis of text layout differences from the global view by measuring the conformance difference between text area and its background. Based on conformance features, our method has rotational invariance, semantic independence, anti-noise, background filtering and multi-scale characteristics and can detect photo-text-mixed images, manuscripts, distorted text cases. The experiments on MARG, Tobacco and MTDD public databases show the effectiveness of this method, and the comparison with five state-of-the-arts methods show the advantages of our method in applicability and effectiveness. 3) We propose a saliency object detection method based on the color conformance operator. We build the conformance operator in Lab color space to measure the conformance of data cube which contains the relations of local data elements. Then we get the saliency map by calculating each small patches of image, and compare their conformance in the global level to obtain the saliency regions of image due to the difference between targets and their background in conformance. Our method can do the saliency object detection by simple classification algorithm without complex feature extraction process and priori information. The experimental results in MSRA-1000 public datasets and its Ground Truth show the effectiveness of our method and the effects in reducing features number and simplifying detection algorithm. 4) We propose a faulty insulator detection method based on repetitive features. The repetitive feature vector is constructed to reflect the structure repeatability of an insulator. The faulty insulator can be detected from the normal ones by calculating the distribution of vector elements. By searching the outliers of a feature vector, we can get the defect position of a faulty insulator. The experimental results in insulator images and UAV videos show the effectiveness, accuracy and robust of our method. Our method can deal with not only multi-defects in one insulator, but also the defects in multi-insulator string. Repeatability is the characterized by conformance reflected in a vector and our method is based on the conformance measurement in one-dimensional. 5) We fuse conformance features with other information aiming to propose some new applicable methods for different research works. We propose three new methods from the conformance ideas. The face image illumination compensation algorithm is proposed based on the illumination conformance and geometric lighting systems modeling, which uses the conformance of a single image to estimating lighting conditions. Combined lighting systems and biometrics knowledge, we propose a new face illumination model with the face image illumination compensation, which can get correct recognition results in the case of large lighting angle conditions in Yale B and extend YaleB(the public data set for face recognition). The gesture recognition algorithm based on depth conformance is proposed, which can measure the conformance differences in depth and get the direction invariance gesture recognition results by combining skin color features and new gestures modeling. The tire-print recognition algorithm based on the uniform conformance is proposed, which uses the conformance inside the tire print and backgrounds to extract the feature points and achieve tire recognition by matching feature points set of candidate species. The experimental results of algorithms presented above show the feasibility and effectiveness of the conformance feature fusion. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 127页 |
| 源URL | [http://ir.sia.ac.cn/handle/173321/16805]  |
| 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 | 蔺蘭. 图像一致性特征与目标检测方法研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2014. |
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。