Research on extracting method of micro-scale remote sensing information combination and application in coastal zone

973 Project of China [2006CB701305]; "863" Project of China [2009AA12Z148]; National Natural Science Foundation of China [40971224]

基于马尔可夫随机场模型的图像分割技术研究

本文针对基于马尔可夫随机场模型（MRF）的图像分割技术进行研究，通过深入分析马尔可夫随机场模型用于图像分割时的优缺点，提出了改进方案，将其用于单帧图像的无监督分割和动态场景下的运动目标分割。主要研究内容包括以下几部分。 第一部分详细介绍了马尔可夫随机场模型，包括邻域系统和基团的概念、初始标记场的获取、能量函数的确立和MAP估算方法。 第二部分针对噪声图像的预处理，提出一种多尺度双边滤波算法来综合不同尺度下双边滤波的去噪效果。为降低双边滤波的计算复杂性，提出一种双边滤波快速计算方法。该算法能够在去除噪声的同时较好地保留边缘。 第三部分针对MRF模型用于图像分割中遇到的过平滑问题，定义了一种间断自适应高斯马尔可夫随机场模型（DA-GMRF），提出一种基于该模型的无监督图像分割方法。利用灰度直方图势函数自动确定分类数及分割阈值，进行多阈值分割得到标记场的初始化，用Metroplis采样器算法进行标记场的优化，得到最终的分割结果。该方法考虑了平滑约束在图像边缘处的自适应性，避免了边缘处的过平滑，将其应用于无监督图像分割取得了较好的效果。 第四部分针对动态场景下的运动目标分割，提出一种基于间断自适应时空马尔可夫随机场模型的运动目标分割方法。解决了传统时空马尔可夫随机场模型不能对运动造成的显露遮挡现象进行处理问题，也克服了全局一致平滑假设造成的过平滑问题。帧差图像二值化得到初始标记场，初始标记场进行‘与’操作获得共同标记场，用Metroplis采样器算法实现共同标记场的优化。该方法既使用了平滑约束，而又保留了间断，从而使分割得到的运动目标边缘更加准确。

曲线演化图像分割和森林烟火预警系统

随着计算机技术，图像采集技术和数据存储技术等的进步，图像处理的应用领域越来越广泛。很多的应用系统是综合利用了电子，通讯和图像处理等技术而开发出来的，图像处理往往是系统的核心部分。图像分割是图像处理的核心技术，也是图像处理技术中的难点。所以研究图像分割技术具有非常重要的意义。 传统的图像分割方法有：使用模板对图像进行边缘检测等；利用滤波处理，频谱分析等数字信号处理处理技术进行分割。80年代末以来，偏微分方程方法越来越多地应用到图像分割领域中，已成为图像分割的有力工具。本文对基于偏微分方程的图像分割方法进行研究，介绍单开曲线演化分割算法，并基于Mumford-Shah模型提出一种带状目标分割方法。这种方法能将图像中的带状区域从图像中分割出来－这里假定带状区域的边界可用单值函数表示。与其它方法，如边缘检测分割，C-V模型分割和单开曲线分割相比，本文提出的方法得到的分割结果有与目标的边界吻合的更好，抗噪能力强等优点。 本文介绍了通过对可见光摄像机所拍摄图像进行分析来检测火的森林烟火预警系统。该系统是通过检测烟的存在来判断是否有火情。图像处理软件是森林烟火预警系统的核心组成部分。评价火灾预警系统性能有两个标准。一个是一旦发生火灾，预警系统能否快速地发出火警信号；另一个是在没有火情时，预警系统是否不报警，即误警率是否低。图像分割在设计图像处理算法时，主要在两个地方得到应用。在图像预处理阶段，利用单开曲线演化分割算法或带状区域的分割算法将森林区域分割出来。这样是为了在对图像进行处理时消除非森林区域中的目标对识别结果的影响，降低误警率。在图像处理阶段，利用图像分割算法将烟从图像中分割出来，准确及时报警。

基于变分的主动轮廓图像分割方法

图像分割是图像处理中很重要的一个问题，是计算机视觉的基础。因为它能够简化信息的存储和表示，从而能够对获取的图像内容进行智能解释，所以在很多应用问题中，图像分割是必不可少的过程，如医学图像处理，环境三维重建及自动目标识别等。图像分割的方法有很多种，如边缘检测，阈值，区域融合，分水岭及马尔可夫随机场等。虽然这些方法有其各自特点，但是它们在图象分割过程中不能充分将图像底层信息与高层信息结合，从而无法模拟人类视觉系统智能性。当图像底层信息不足时，这些仅基于数据驱动的分割模型无法达到令人满意的结果。尽管某种具体图像分割方法不可能满足所有图像分割要求，但利用尽可能多的高层与底层信息，将图像分割成有意义和人们所期望的区域始终是研究者所追求的目标。图像分割问题的数学建模和计算中有两个关键因素。第一是建立合适的分割模型将分割边界和分割区域的作用有效结合。第二是利用最有效的方法将分割边界和分割区域的几何特征统一到分割模型中。基于变分原理的主动轮廓图像分割将图像视为连续函数。这就使得研究者可以从连续函数空间角度来研究图像分割问题。这同时也为研究者提供严格的数学工具，如微分几何、泛函分析和微分方程等。为此它能很好的解决上述两个问题。第一，Mumford-Shah(M-S)模型为基于变分的主动轮廓分割模型提供了一完整的数学理论框架，并且Mumford-Shah模型从信息论的角度也能得到合理解释。第二，水平集方法能有效的表示分割边界和分割区域的几何特征。与其它方法相比，变分主动轮廓在理论和实际计算过程中都具有显著的优势。首先它能直接处理和表示各种重要的几何特征，如梯度、切向量、曲率等，并且能有效模拟很多动态过程，如线性和非线性扩散等。再则其可以利用很多已有的丰富数值方法进行分析和计算。本文基于变分原理与偏微分方程方法，利用主动轮廓模型具有结合底层图像信息与高层先验知识的特点，将特定先验知识与主动轮廓分割模型进行有效结合以弥补底层图像信息的不足，从而使主动轮分割廓模型具有更强的智能性。本文主要从两点对变分主动轮廓分割模型展开了研究：1、演化轮廓的形状约束；2、演化轮廓的梯度下降流约束及其滤波实现。其主要工作包括以下四个方面的内容：第一，基于M-S模型和样条曲线的开边界检测。本章通过对演化轮廓引入合理边界条件，利用样条曲线表示待检测的开曲线，将一般开曲线的检测问题变为合理的图像分割问题，从而达到一般开曲线检测目的。此方法称为开扩散蛇模型。一般开曲线的检测具有很多应用领域，如：河流、道路、天际线、焊缝等检测。第二，方差主动轮廓模型。在目标跟踪应用中，跟踪目标的主要运动形式表现为平移。本章将此作为一种先验知识与主动轮廓模型结合，提出了一种方差主动轮廓模型(HV)。此模型的特点是轮廓在演化过程中具有平移优先和快速的良好特性。它比已有的主动轮廓模型更适于自动目标跟踪领域。第三，基于M-S模型和隐式曲面变分方法的一般梯度下降流滤波器。本章为一般梯度下降流求取提供了统一框架及解决方法。首先本章将H0梯度下降流和一般梯度下降流统一到Mumford-Shah模型框架中，从而将一般梯度下降流的求取转换为一个极小化泛函问题，并利用隐式曲面变分方法对此极小化泛函进行求解。另外本章从滤波器设计角度出发，通过对H0梯度下降流滤波可以得到一般梯度下降流。滤波器的实现体现了内嵌于一般梯度下降流的先验属性。根据此思想，本章将对应于HV和H1主动轮廓的內积空间顺序组合，对H0梯度下降流进行顺序滤波，提出了一种既具有全局平移优先性又具有局部光滑速度场的主动轮廓，称为HV1主动轮廓。它将H0，H1和HV主动轮廓统一起来。第四，形状保持主动轮廓模型及其应用。针对某些特定目标的检测，本章提出了形状保持主动轮廓模型。此模型能够达到分割即目标的目的，同时能够给出目标的定量描述。基于此模型，本章实现了具有椭圆、直线和平行四边形轮廓特征目标的检测。椭圆形状约束用于眼底图像分割。直线和平行四边行分别用于自动目标识别中的天际线检测和机场跑道跟踪。

动态有序变化的阈值曲面方法

对于背景灰度不一致的图像的二值化问题,可采用在图像的不同区域选取不同阈值的动态阈值曲面方法解决。本文提出一种有序变化的动态阈值曲面方法,通过选取三个动态变化调节系数,使曲面动态地逼近阈值分割的最佳值,同时又降低曲面的误分割,能较好地提取二值图像。

一种基于知识的自然景物解释

利用我们研制成的基于知识的区域分析系统,我们对自然景物图象做了分析、解释。本文介绍:(1)图象初始分割;(2)图象特征提取;(3)规则集构造。这里,图象初始分割采用的是模糊域方法,它基于Fuzzy C-means 算法,并在此基础上修改了收敛准则,增加了迭代分割功能。图象特征分为主持征及从特征,它们建立在层次化的区域数据结构上。适合于区域分析的规则集已包括三类规则,它们不仅具有较强的知识表示能力,而且易于控制及利用。本文介绍针对包括天空、道路、树木、建筑物等物体的简单景物图象所做的解释实验,并给出了实验结果.

Optical determination of fractal dimensions using Fourier transforms

R. Zwiggelaar and C.R. Bull, 'Optical determination of fractal dimensions using Fourier transforms', Optical Engineering 34 (5), 1325-1332 (1995)

MosaicShape: Stochastic Region Grouping with Shape Prior

A novel method that combines shape-based object recognition and image segmentation is proposed for shape retrieval from images. Given a shape prior represented in a multi-scale curvature form, the proposed method identifies the target objects in images by grouping oversegmented image regions. The problem is formulated in a unified probabilistic framework and solved by a stochastic Markov Chain Monte Carlo (MCMC) mechanism. By this means, object segmentation and recognition are accomplished simultaneously. Within each sampling move during the simulation process,probabilistic region grouping operations are influenced by both the image information and the shape similarity constraint. The latter constraint is measured by a partial shape matching process. A generalized parallel algorithm by Barbu and Zhu,combined with a large sampling jump and other implementation improvements, greatly speeds up the overall stochastic process. The proposed method supports the segmentation and recognition of multiple occluded objects in images. Experimental results are provided for both synthetic and real images.

Texture Segregation By Visual Cortex: Perceptual Grouping, Attention, and Learning

A neural model is proposed of how laminar interactions in the visual cortex may learn and recognize object texture and form boundaries. The model brings together five interacting processes: region-based texture classification, contour-based boundary grouping, surface filling-in, spatial attention, and object attention. The model shows how form boundaries can determine regions in which surface filling-in occurs; how surface filling-in interacts with spatial attention to generate a form-fitting distribution of spatial attention, or attentional shroud; how the strongest shroud can inhibit weaker shrouds; and how the winning shroud regulates learning of texture categories, and thus the allocation of object attention. The model can discriminate abutted textures with blurred boundaries and is sensitive to texture boundary attributes like discontinuities in orientation and texture flow curvature as well as to relative orientations of texture elements. The model quantitatively fits a large set of human psychophysical data on orientation-based textures. Object boundar output of the model is compared to computer vision algorithms using a set of human segmented photographic images. The model classifies textures and suppresses noise using a multiple scale oriented filterbank and a distributed Adaptive Resonance Theory (dART) classifier. The matched signal between the bottom-up texture inputs and top-down learned texture categories is utilized by oriented competitive and cooperative grouping processes to generate texture boundaries that control surface filling-in and spatial attention. Topdown modulatory attentional feedback from boundary and surface representations to early filtering stages results in enhanced texture boundaries and more efficient learning of texture within attended surface regions. Surface-based attention also provides a self-supervising training signal for learning new textures. Importance of the surface-based attentional feedback in texture learning and classification is tested using a set of textured images from the Brodatz micro-texture album. Benchmark studies vary from 95.1% to 98.6% with attention, and from 90.6% to 93.2% without attention.

Micro-Anatomical Quantitative Imaging Towards Enabling Automated Diagnosis of Thick Tissues at the Point of Care

Histopathology is the clinical standard for tissue diagnosis. However, histopathology has several limitations including that it requires tissue processing, which can take 30 minutes or more, and requires a highly trained pathologist to diagnose the tissue. Additionally, the diagnosis is qualitative, and the lack of quantitation leads to possible observer-specific diagnosis. Taken together, it is difficult to diagnose tissue at the point of care using histopathology.

Several clinical situations could benefit from more rapid and automated histological processing, which could reduce the time and the number of steps required between obtaining a fresh tissue specimen and rendering a diagnosis. For example, there is need for rapid detection of residual cancer on the surface of tumor resection specimens during excisional surgeries, which is known as intraoperative tumor margin assessment. Additionally, rapid assessment of biopsy specimens at the point-of-care could enable clinicians to confirm that a suspicious lesion is successfully sampled, thus preventing an unnecessary repeat biopsy procedure. Rapid and low cost histological processing could also be potentially useful in settings lacking the human resources and equipment necessary to perform standard histologic assessment. Lastly, automated interpretation of tissue samples could potentially reduce inter-observer error, particularly in the diagnosis of borderline lesions.

To address these needs, high quality microscopic images of the tissue must be obtained in rapid timeframes, in order for a pathologic assessment to be useful for guiding the intervention. Optical microscopy is a powerful technique to obtain high-resolution images of tissue morphology in real-time at the point of care, without the need for tissue processing. In particular, a number of groups have combined fluorescence microscopy with vital fluorescent stains to visualize micro-anatomical features of thick (i.e. unsectioned or unprocessed) tissue. However, robust methods for segmentation and quantitative analysis of heterogeneous images are essential to enable automated diagnosis. Thus, the goal of this work was to obtain high resolution imaging of tissue morphology through employing fluorescence microscopy and vital fluorescent stains and to develop a quantitative strategy to segment and quantify tissue features in heterogeneous images, such as nuclei and the surrounding stroma, which will enable automated diagnosis of thick tissues.

To achieve these goals, three specific aims were proposed. The first aim was to develop an image processing method that can differentiate nuclei from background tissue heterogeneity and enable automated diagnosis of thick tissue at the point of care. A computational technique called sparse component analysis (SCA) was adapted to isolate features of interest, such as nuclei, from the background. SCA has been used previously in the image processing community for image compression, enhancement, and restoration, but has never been applied to separate distinct tissue types in a heterogeneous image. In combination with a high resolution fluorescence microendoscope (HRME) and a contrast agent acriflavine, the utility of this technique was demonstrated through imaging preclinical sarcoma tumor margins. Acriflavine localizes to the nuclei of cells where it reversibly associates with RNA and DNA. Additionally, acriflavine shows some affinity for collagen and muscle. SCA was adapted to isolate acriflavine positive features or APFs (which correspond to RNA and DNA) from background tissue heterogeneity. The circle transform (CT) was applied to the SCA output to quantify the size and density of overlapping APFs. The sensitivity of the SCA+CT approach to variations in APF size, density and background heterogeneity was demonstrated through simulations. Specifically, SCA+CT achieved the lowest errors for higher contrast ratios and larger APF sizes. When applied to tissue images of excised sarcoma margins, SCA+CT correctly isolated APFs and showed consistently increased density in tumor and tumor + muscle images compared to images containing muscle. Next, variables were quantified from images of resected primary sarcomas and used to optimize a multivariate model. The sensitivity and specificity for differentiating positive from negative ex vivo resected tumor margins was 82% and 75%. The utility of this approach was further tested by imaging the in vivo tumor cavities from 34 mice after resection of a sarcoma with local recurrence as a bench mark. When applied prospectively to images from the tumor cavity, the sensitivity and specificity for differentiating local recurrence was 78% and 82%. The results indicate that SCA+CT can accurately delineate APFs in heterogeneous tissue, which is essential to enable automated and rapid surveillance of tissue pathology.

Two primary challenges were identified in the work in aim 1. First, while SCA can be used to isolate features, such as APFs, from heterogeneous images, its performance is limited by the contrast between APFs and the background. Second, while it is feasible to create mosaics by scanning a sarcoma tumor bed in a mouse, which is on the order of 3-7 mm in any one dimension, it is not feasible to evaluate an entire human surgical margin. Thus, improvements to the microscopic imaging system were made to (1) improve image contrast through rejecting out-of-focus background fluorescence and to (2) increase the field of view (FOV) while maintaining the sub-cellular resolution needed for delineation of nuclei. To address these challenges, a technique called structured illumination microscopy (SIM) was employed in which the entire FOV is illuminated with a defined spatial pattern rather than scanning a focal spot, such as in confocal microscopy.

Thus, the second aim was to improve image contrast and increase the FOV through employing wide-field, non-contact structured illumination microscopy and optimize the segmentation algorithm for new imaging modality. Both image contrast and FOV were increased through the development of a wide-field fluorescence SIM system. Clear improvement in image contrast was seen in structured illumination images compared to uniform illumination images. Additionally, the FOV is over 13X larger than the fluorescence microendoscope used in aim 1. Initial segmentation results of SIM images revealed that SCA is unable to segment large numbers of APFs in the tumor images. Because the FOV of the SIM system is over 13X larger than the FOV of the fluorescence microendoscope, dense collections of APFs commonly seen in tumor images could no longer be sparsely represented, and the fundamental sparsity assumption associated with SCA was no longer met. Thus, an algorithm called maximally stable extremal regions (MSER) was investigated as an alternative approach for APF segmentation in SIM images. MSER was able to accurately segment large numbers of APFs in SIM images of tumor tissue. In addition to optimizing MSER for SIM image segmentation, an optimal frequency of the illumination pattern used in SIM was carefully selected because the image signal to noise ratio (SNR) is dependent on the grid frequency. A grid frequency of 31.7 mm-1 led to the highest SNR and lowest percent error associated with MSER segmentation.

Once MSER was optimized for SIM image segmentation and the optimal grid frequency was selected, a quantitative model was developed to diagnose mouse sarcoma tumor margins that were imaged ex vivo with SIM. Tumor margins were stained with acridine orange (AO) in aim 2 because AO was found to stain the sarcoma tissue more brightly than acriflavine. Both acriflavine and AO are intravital dyes, which have been shown to stain nuclei, skeletal muscle, and collagenous stroma. A tissue-type classification model was developed to differentiate localized regions (75x75 µm) of tumor from skeletal muscle and adipose tissue based on the MSER segmentation output. Specifically, a logistic regression model was used to classify each localized region. The logistic regression model yielded an output in terms of probability (0-100%) that tumor was located within each 75x75 µm region. The model performance was tested using a receiver operator characteristic (ROC) curve analysis that revealed 77% sensitivity and 81% specificity. For margin classification, the whole margin image was divided into localized regions and this tissue-type classification model was applied. In a subset of 6 margins (3 negative, 3 positive), it was shown that with a tumor probability threshold of 50%, 8% of all regions from negative margins exceeded this threshold, while over 17% of all regions exceeded the threshold in the positive margins. Thus, 8% of regions in negative margins were considered false positives. These false positive regions are likely due to the high density of APFs present in normal tissues, which clearly demonstrates a challenge in implementing this automatic algorithm based on AO staining alone.

Thus, the third aim was to improve the specificity of the diagnostic model through leveraging other sources of contrast. Modifications were made to the SIM system to enable fluorescence imaging at a variety of wavelengths. Specifically, the SIM system was modified to enabling imaging of red fluorescent protein (RFP) expressing sarcomas, which were used to delineate the location of tumor cells within each image. Initial analysis of AO stained panels confirmed that there was room for improvement in tumor detection, particularly in regards to false positive regions that were negative for RFP. One approach for improving the specificity of the diagnostic model was to investigate using a fluorophore that was more specific to staining tumor. Specifically, tetracycline was selected because it appeared to specifically stain freshly excised tumor tissue in a matter of minutes, and was non-toxic and stable in solution. Results indicated that tetracycline staining has promise for increasing the specificity of tumor detection in SIM images of a preclinical sarcoma model and further investigation is warranted.

In conclusion, this work presents the development of a combination of tools that is capable of automated segmentation and quantification of micro-anatomical images of thick tissue. When compared to the fluorescence microendoscope, wide-field multispectral fluorescence SIM imaging provided improved image contrast, a larger FOV with comparable resolution, and the ability to image a variety of fluorophores. MSER was an appropriate and rapid approach to segment dense collections of APFs from wide-field SIM images. Variables that reflect the morphology of the tissue, such as the density, size, and shape of nuclei and nucleoli, can be used to automatically diagnose SIM images. The clinical utility of SIM imaging and MSER segmentation to detect microscopic residual disease has been demonstrated by imaging excised preclinical sarcoma margins. Ultimately, this work demonstrates that fluorescence imaging of tissue micro-anatomy combined with a specialized algorithm for delineation and quantification of features is a means for rapid, non-destructive and automated detection of microscopic disease, which could improve cancer management in a variety of clinical scenarios.

Observational Evidence for Mode Coupling in the Chromospheric Network

Oscillations in network bright points (NBPs) are studied at a variety of chromospheric heights. In particular, the three-dimensional variation of NBP oscillations is studied using image segmentation and cross-correlation analysis between images taken in light of Ca II K3, Ha core, Mg I b2, and Mg I b1-0.4 Å. Wavelet analysis is used to isolate wave packets in time and to search for height-dependent time delays that result from upward- or downward-directed traveling waves. In each NBP studied, we find evidence for kink-mode waves (1.3, 1.9 mHz), traveling up through the chromosphere and coupling with sausage-mode waves (2.6, 3.8 mHz). This provides a means for depositing energy in the upper chromosphere. We also find evidence for other upward- and downward-propagating waves in the 1.3-4.6 mHz range. Some oscillations do not correspond to traveling waves, and we attribute these to waves generated in neighboring regions.

An FPGA Based Coprocessor for GLCM and Haralick Texture Features and Their Application in Prostate Cancer Classification

Grey Level Co-occurrence Matrix (GLCM), one of the best known tool for texture analysis, estimates image properties related to second-order statistics. These image properties commonly known as Haralick texture features can be used for image classification, image segmentation, and remote sensing applications. However, their computations are highly intensive especially for very large images such as medical ones. Therefore, methods to accelerate their computations are highly desired. This paper proposes the use of programmable hardware to accelerate the calculation of GLCM and Haralick texture features. Further, as an example of the speedup offered by programmable logic, a multispectral computer vision system for automatic diagnosis of prostatic cancer has been implemented. The performance is then compared against a microprocessor based solution.

Semi-Qualitative Probabilistic Networks in Computer Vision Problems

This paper explores the application of semi-qualitative probabilistic networks (SQPNs) that combine numeric and qualitative information to computer vision problems. Our version of SQPN allows qualitative influences and imprecise probability measures using intervals. We describe an Imprecise Dirichlet model for parameter learning and an iterative algorithm for evaluating posterior probabilities, maximum a posteriori and most probable explanations. Experiments on facial expression recognition and image segmentation problems are performed using real data.

Anomaly Detection for Data with Spatial Attributes

The problem of detecting spatially-coherent groups of data that exhibit anomalous behavior has started to attract attention due to applications across areas such as epidemic analysis and weather forecasting. Earlier efforts from the data mining community have largely focused on finding outliers, individual data objects that display deviant behavior. Such point-based methods are not easy to extend to find groups of data that exhibit anomalous behavior. Scan Statistics are methods from the statistics community that have considered the problem of identifying regions where data objects exhibit a behavior that is atypical of the general dataset. The spatial scan statistic and methods that build upon it mostly adopt the framework of defining a character for regions (e.g., circular or elliptical) of objects and repeatedly sampling regions of such character followed by applying a statistical test for anomaly detection. In the past decade, there have been efforts from the statistics community to enhance efficiency of scan statstics as well as to enable discovery of arbitrarily shaped anomalous regions. On the other hand, the data mining community has started to look at determining anomalous regions that have behavior divergent from their neighborhood.In this chapter,we survey the space of techniques for detecting anomalous regions on spatial data from across the data mining and statistics communities while outlining connections to well-studied problems in clustering and image segmentation. We analyze the techniques systematically by categorizing them appropriately to provide a structured birds eye view of the work on anomalous region detection;we hope that this would encourage better cross-pollination of ideas across communities to help advance the frontier in anomaly detection.