目标成像计算与模拟算法研究
文献类型:学位论文
作者 | 吴登禄![]() |
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 | 目标成像 成像建模 光谱反射率 成像公式 虚拟成像计算 |
其他题名 | Research on Object Imaging Computation and Simulation |
学位专业 | 模式识别与智能系统 |
中文摘要 | 目标成像与目标跟踪都是计算机视觉中的基本任务。目标成像是目标跟踪的基础,也是图像分析的前提和基础。Marr提出了计算机视觉理论,将计算机视觉分为底层视觉、中层视觉及高层视觉,这三个不同层次的视觉之间是一种自下而上的依赖的关系。底层视觉处理图像的像素信息和颜色信息以及对在此基础上形成的图像特征进行抽取和分析,如:角点提取、边缘检测、纹理信息等。然而,图像的中层分析和高层理解都是以上述的图像特征为基础对图像进行理解和分析,因此目标成像就成为计算机视觉领域研究的基础问题。 图像的颜色和亮度等信息在成像过程中会受到环境中各种因素的影响。这些因素可以分为外部影响因素和内部影响因素。在目标成像过程中,相机外部环境因素主要包括光源、目标物体的光谱反射率;内部影响因素主要是相机内部对响应值的各种处理算法的影响。因此,对目标成像过程中的各种影响因素进行分析,将会提高各种计算机视觉算法的鲁棒性,促进计算机视觉的发展。 本文通过对上述影响因素的分析和建模,提出了目标物体的光谱反射率、相机的光谱响应曲线计算方法,并在上述计算基础上实现了相机模拟成像,同时提出了新的成像模型。主要的创新工作内容及成果有如下五部分: 1、目标成像因素的分析:本论文对影响相机成像的因素:光源、目标的光谱反射率及相机本身的响应特性和成像后处理算法进行了分析。同时,通过分析整个成像过程,提出把相机的成像过程分为两个阶段:Out-camera阶段和In-camera阶段,并对相机成像的上述两个过程进行了建模。 2、相机光谱响应曲线的计算方法:相机光谱响应曲线是相机本身的特性,同时也是相机成像过程中重要的影响因素。由于相机光谱响应曲线在实际中不能很容易的获得,因此本文针对上述问题,并通过对相机成像过程和对相机光谱响应曲线特性的分析,提出了一种新的基于RAW数据的相机光谱响应曲线计算方法,提高了相机光谱响应曲线计算的精度。 3、目标光谱反射率基函数的计算:目标光谱反射率作为目标物体本身的一种特性,被称为物体的“指纹”。由于自然界中物体的光谱反射率可以通过基函数近似线性表示,因此基函数的精度将对目标物体光谱反射率的恢复带来很大的影响。本文针对上述问题,提出了一种新的基于优化的光谱反射率基函数计算方法,通过实验证明其结果优于经典的PCA方法所获得的基函数。 4、目标物体的虚拟成像及新的成像公式:在第一章对影响相机成像因素进行分析的基础上,后面章节又分别对上述影响因素进行了定量的计算。本章在前面几章的基础上,通过成像公式对相机成像的两阶段模型进行了成像模拟计算,为相机的设计以及成像因素对成像的定量化分析提供了重要的基础。在上述成像计算的基础上提出了新的成像公式,并利用新的成像公式提出了新的目标场景复现的方法。 5、目标成像计算及模拟应用——本章通过总结基于图像的目标识别方法,并以具体的基于图像的目标识别应用为例,通过前面几章提出的成像模拟计算的方法,对实际场景进行了成像模拟,分析了基于图像目标识别方法在实际问题中的缺点,并结合前面第二章从图像中反解出的目标物体光谱反射率,提出了把基于图像的目标识别方法和目标物体的光谱信息进行融合,以提高目标识别的准确率和鲁棒性。 |
索取号 | TP391.41/W81/2014 |
英文摘要 | Object imaging and object recognition are fundamental tasks in computer vision field, as important contents of image processing and premise of image analysis and image understanding. Marr proposed to understand vision at three different levels: low level vision, middle level vision and high level vision. The relationships are top-down among the three different levels. The low level vision aims to extract the color and brightness information and derivated other features from the image, such as: corner, edge and texture. While the middle and high level visions are based on the above information for analysis and understanding the image. Therefore, object imaging has also become the fundamental problem for computer vision. The color, brightness and derivated features have impact on the stability and robustness of the algorithms. The color and brightness information about the image are influenced by different factors in the imaging process. These factors include the environment and the post processing algorithm about the image. The environment factors are composed by the light source character, the reflectance spectra of the object and the sensor spectral sensitivity. The post processing factors about the image include the different algorithms for image processing inside the camera. Therefore, analyzing the different factors improve the robustness and stability of the algorithms. The main research direction in this dissertation is to modeling, analyzing and computation for the impact factors. We proposed some new methods for the sensor spectral sensitivity function, the object reflectance spectra. Based on the above analyzing and modeling, we realized the object imaging simulation and proposed a new imaging model. The main contents and contributions are summarized as follows: 1) The impact factors analysis about the object imaging: We analyze the impact factors about the object imaging including the illumination of the light source, the reflectance spectra of the object, the sensor spectral sensitivity and the post processing algorithms of the image. Based on the above analysis, we divided the object imaging process into two different steps: the out-camera imaging and in-camera imaging. 2) The computation of the sensor spectral sensitivity: The sensor spectral sensitivity is the character of the camera and an important factor influenced the object imaging. In practice, the sensor spectral sensitivity is not easily obtained. Therefore, we proposed a new method based on the RAW data for solving the problem. The experiments show that our method is better than the state-art method. 3) The basis function of the object reflectance spectra: The reflectance spectra is a character of the object called the “finger print”. Due to the fact that the natural reflectance can approximated by several basis functions, the accuracy of the basis function can bring the huge impact of the reflectance spectra reconstruction. In this chapter, we proposed a new method based on the optimization. The experiment validates that the results are more accurate than the PCA method. 4) The object imaging simulation and a new imaging model: Based on the previous analysis and making a quantity computation for the impact factors, we simulated the camera imaging according to the out-camera model and in-camera model. Through the computation and analysis, we proposed a new imaging model. According to the new imaging model, we also proposed a new method for the scene reproduction. 5) The application of the object imaging computation and simulation: This chapter summarize the object recognition method based on the image. Through the above method proposed imaging computation and simulation, we simulate the imaging process in practice and analysis the disadvantage of the recognition method based on the image. Meanwhile, according to the method which calculates the reflectance spectra from the image, we proposed a new method that fused the image and reflectance spectra information for object recognition in order to improve the accuracy and robustness of the object recognition algorithms. |
语种 | 中文 |
产权排序 | 1 |
页码 | 109页 |
源URL | [http://ir.sia.ac.cn/handle/173321/16801] ![]() |
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 | 吴登禄. 目标成像计算与模拟算法研究[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。