焦探测器和一个倒置显微镜（尼康TE200 ） [ 15 ] 。多光子显微镜可以准确地找到一个三维荧光量，可成功地应用于分析血管形态。通常是一个小肿瘤（直径几毫米）植入皮肤的窗口中庭。整个肿瘤血管的影像大多数实验。图片10倍的目标而采取的一切，但最小的 肿瘤和图像覆盖约1.3x1.3毫米的组织。成堆的图像所采取的一个典型的堆栈50片。它通常需要13分钟获得的图像为整个堆栈 3 。图像处理-方法
主要有两个阶段在处理这些图像： 1 ）图像拼接
拼接是由滑动的新形象的综合形象和寻找最佳关联点。 2 ）图像融合
配煤是由分离颜色的飞机，在必要情况下，采用混合算法每个彩色带 重组飞机一起获得全彩色图像的输出。混合图像应保持质量输入图像[ 16 ] 。这些过程中有详细的解释，并参阅下文二维图像，除非明确指出，他们提到的三维图像。算法开发了C编程语言LabWindows / CVI的7.0 （美国国家仪器有限公司）开发环境，使用IMAQ图像处理图书馆和Windows XP专业版操作系统。那个算法是完全自动的，他们已经在电脑上测试的处理器速度1.53GHz和448MB的内存。 3.1拼接方法
在该算法的缝合是由图像翻译只。应用程序可以被称为作为拼接，瓦工， montaging或缝合。第一步是生成的相对位置所获得的图像和建立一个空的图像阵列在电脑记忆体，这些图片将放在。下一步是搜索对于这一点的最佳关联是由相邻的图像边缘滑动是双向的，直到最佳比赛的边缘特征发现。这个搜索过程需要选择最佳的搜索空间如图1所示，在其中进行搜索的最佳关联。使用太多像素内使这个方块相关过程耗时太少像素，同时减少比赛的质量。选择若干像素使用密切相关的各个方面的功能预期将显着的形象而这又取决于重点质量，即对目前的最大空间频率的形象。
附件2：外文原文
An algorithm for image stitching and blending ABSTRACT
In many clinical studies, including those of cancer, it is highly desirable to acquire images of whole tumour sectionswhilst retaining a microscopic resolution. A usual approach to this is to create a composite image by appropriatelyoverlapping individual images acquired at high magnification under a microscope. A mosaic of these images can beaccurately formed by applying image registration, overlap removal and blending techniques. We describe an optimised,automated, fast and reliable method for both image joining and blending. These algorithms can be applied to most typesof light microscopy imaging. Examples from histology, from in vivo vascular imaging and from fluorescenceapplications are shown, both in 2D and 3D. The algorithms are robust to the varying image overlap of a manually moved stage, though examples of composite images acquired both with manually-driven and computer-controlled stages are presented. The overlap-removal algorithm is based on the cross-correlation method; this is used to determine and select the best correlation point between any new image and the previous composite image. A complementary image blending algorithm, based on a gradient method, is used to eliminate sharp intensity changes at the image joins, thus gradually blending one image onto the adjacent ‘composite’. The details of the algorithm to overcome both intensity discrepancies and geometric misalignments between the stitched images will be presented and illustrated with several examples.
Keywords: Image Stitching, Blending, Mosaic images 1. INTRODUCTION AND BACKGROUND
There are many applications which require high resolution images. In bright-field or epifluorescence microscopy [1],for example, which are used in biological and medical applications, it is often necessary to analyse
a complete tissue section which has dimensions of several tens of millimetres, at high resolution. However, the high resolution single image cannot be realised with a low power objective, necessary to view a large sample, even if using cameras with tens of millions of active pixels. The most common approach is to acquire several images of parts of the tissue at high magnification and assemble them into a composite single image which preserves the high resolution. This process of assembling the composite image from a number of images, also known as ‘tiling’ or ‘mosaicing’ requires an algorithm for image stitching (registration) and blending. The automatic creation of large high resolution image mosaics is a growing research area involving computer vision and image processing. Mosaicing with blending can be defined as producing a single edgeless image by putting together a set of overlapped images [2]. Automating this process is an important issue as it is difficult and time consuming to achieve it manually. One such algorithm for image stitching and blending is presented in this paper. Image stitching combines a number of images taken at high resolution into a composite image. The composite image must consist of images placed at the right position and the aim is to make the edges between images invisible. The quality of stitching is therefore expressed by measuring both the correspondence between adjacent stitched images that form the composite image and the visibility of the seam between the stitched images [3]. Image stitching (registration) methods have been explained in detail in [4]. In [5], cross-correlation is shown to be the preferred method for automatic registration of large number of images. Various registration methods were compared in this paper [5] and it was showed that the cross-correlation method provided the smallest error. When these methods were compared in terms of speed, the cross-correlation was shown to be the second fastest but much more accurate than the fastest method (principal axes method). There are a number of papers that deal with the stitching problem [3, 6-8]. Image stitching can be performed using image pixels directly - correlation method;
in frequency domain - fast Fourier transform method; using low level Three-Dimensional and Multidimensional Microscopy: Image Acquisition and Processing XII, Jose-Angel Conchello, Carol J. Cogswell, Tony Wilson, Editors, March 2005 191 features such as edges and corners; using high level features such as parts of objects [2]. Brown [4] classifies image registration according to following criteria: type of feature space, type of search strategies and type of similarity measure.Approaches for image stitching that optimise the search for the best correlation point by using Levenberg-Marquardt method are given in [2, 9, 10]. Levenberg-Marquardt method gives good results, but it is computationally expensive and can get stuck at local minima. An alternative way is to apply an algorithm which searches for the best correlation point by employing a ‘coarse to fine’ resolution approach in order to reduce the number of calculations [10, 11]. The approach offered in this paper makes the selection of the best correlation point in the following way. Based on knowledge about the expected overlap when using the motorised stage, it would be straightforward to find the best correlation point in the ideal case. However, the overlap area is not perfect, and certainly not to an accuracy of one pixel, due to deviations in stage position from the ideal and due to stage/camera misalignment. Our algorithm offers away to overcome this problem by searching the small area around the expected central overlap pixel in order to find the best correlation point. Positioning of acquired images with a manual stage is much less accurate, so there is a need to search a wider area in order to find the best cross-correlation point. Most of the existing methods of image stitching either produce a ‘rough’ stitch that cannot deal with common features such as blood vessels, comet cells and histology, or they require some user input [12]. The new algorithm presented in this paper has embedded code to deal with such features. In order to remove the edges and make one compact image it is necessary to apply additional image blending. The process of image blending is restricted to zones of overlap which are
determined during the stitching process. This means that if the overlap regions between images are large, and images are not perfectly matched on these parts, ghosting or ‘blurring’ is visible. However, if these regions are small, the seams will be visible [13]. In order to avoid these effects and make the blurring effect negligible, the cross-correlation function between the composite image and the image which is to be stitched needs to be applied appropriately. The new method presented in this paper shows that the best quality image can be achieved if blending is applied after each image has been stitched. This approach improves the stitching of additional images because the cross-correlation is applied to a blended composite image which gives a more robust result. When acquiring images of highly non-uniform samples, as it is the case in our in vivo studies, the lighting conditions change and thus influence the cross-correlation applied during stitching. These lighting changes prevent the removal of artefacts. In order to avoid this effect it may be possible to normalise the illumination of the images, but it could cause some loss of information as one cannot be sure what the real cause for the variation in the image illumination is. It can come from the changes in the lighting but also from the different tissue colour. Hence, some illumination compensation is necessary. Our achievement is a high-quality, automatic stitching and blending algorithm that responds to features such as blood vessels, comet cells and histology samples. The illumination compensation is not incorporated in the presented algorithm. This paper is organised as follows. Section 2 explains the image acquisition process. Section 3 explains the methodology followed during the development of the image processing algorithm that applies both the stitching and blending. Section 4 gives the results of the applied algorithm on the selected images after the stitching only and after both stitching and blending and illustrates the effectiveness of the proposed algorithm. Conclusions are presented in Section 5 and directions for the future work are defined.