【作者】 王选贺；

【导师】 陈贺新；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2007， 硕士

【摘要】 尿沉渣显微镜检查是临床检验和诊断鉴别的重要方法。尿沉渣显微图像处理过程,包括图像滤波去除噪声、图像的边缘检测、图像二值化、图像分割、特征提取和分类器设计等。本文在深入研究和大量实验的基础上,对图像的分割、特征提取、分类器设计方面提出一些行之有效的算法。在图像分割方面,对图像进行了抽取压缩,使图像分割效率大大提高,具有较强的应用性。在分割过程中,出现了将一个目标分割成多个区域的结果,本论文发明了一种打“补丁”的算法,使同一个目标合并在一起而不影响其它目标。在特征提取方面,创造性的将图像的旋转、快速傅里叶变换、椭圆拟合与几何形状相结合,对其进行数字化,用于提取细胞的特征,具有较强的区分能力和抗干扰性。在分类器设计方面,采用基于二叉树的神经网络分类系统,二叉树将类似的目标进行粗分类,然后利用神经网络自我学习的能力,通过不断的学习,积累判别的知识,对一些模棱两可的尿沉渣有形成分做出正确的判断,提高了有形成分的识别的准确性。更多还原

【Abstract】 1. Introduction.The examination of urinary sediments is a basic requirement in the regulation of examination and operation in the center of clinic examination .According to shape and texture of urinary sediment ,the doctor identify varieties of physiological or pathological ingredients such as erythrocyte, leukocyte, cast ,epithelial cell, crystal bacterium, parasite and so on .Consequently, the microscope help doctor make a correct diagnose on the diseases of urinary system。In general,if you can’t find the disease in conventional examination and chemical experiment, you’d better analyze material ingredients in urinary sediment by means of quantifications or qualities. Maybe you can get some important information. It will be seen from this that the examination of urinary sediments with microscope has very important clinical significance.2. The Researching Contents.Firstly, the background and the studying meaning of subject are stated in brief in this dissertation.The developing condition of recognition on automatical urine sediments facility in inland and overseas ,the studying content and the key technology of this dissertation are refered.Secondly, the edge extraction and segmentation technology on urine sediment microscopic images are expounded. In the course of statements on segmentation algorithm ,the way in which theories and practices are united are applied.Namely,after introducing segmentation algorithm,that algorithm are applied in urine images in order to understand this algorithm.Thirdly,On top of segmentation,the extracted features of urine sedments microscopic images are expatiated. At first ,we extract the feature of small cells.The features of the rotundity ratio ,rotation squre,the frequence feature on FFT are extracted after prepared processes. Then we extract the feature of large cells.The features of the moment and imitation of ellipse are extracted.The structure of extraction feature refers the arithemetic of feature as main streams.Then feature datas of process on algorithm just now mentioned show by means of scatter picture in order to contrast the different cell.Forthly, this dissertation applies neural net system on bin-tree to identify the cells. At first, the similarity targets are cursorily assorted by means of bin-tree. Then neural net system is used to accurate identification.This algorithm makes good use of the merits of bin-tree and neural net system。At the same time ,this arithemetic reduces the learning time of eural net system.At Last,the frame of overall systems and the process of program on C++ are introduced.During the course of research on microscope image of urinary sediments which are non-centrifugal, there are mainly some innovations in this dissertation as follows:Firstly, in the respect of image segmentation:The segmentations of microscopic image of urinary sediments distinguish between material ingredients in urinary sediment and background, and then cut the material ingredients from the original image.The segmentation of microscopic images is the first step, so it directly affects the digital features and the results of recognition. The segmentation algorithm of this dissertation is as follows. Firstly, I compress the original image by means of extractive pixels, namely, get the even pixels in the direction of row or column, irrespectively, from the original grey image of 800 by 600, and make a new grey image of 400 by 300 which doesn’t affect the segmentation, because there are the correlation and no great leaps between mutual neighboring pixels. Secondly, I extract the edges to the image which was condensed with the edge algorithm of SOBEL. Thirdly, the edge image is transformed into binary image which is a method that if pixel value of the edge image is more than a certain value, we will assign one to a new image, or else we will assign zero to a new image, all the pixel values do like this one by one. The certain value is experimental value. I take two values, namely, we will get two binary images. The one which is a high threshold is to get small cells such as erythrocyte, leukocyte , etc, and the one which is a low threshold is to get big cells such as cast , epithelial cell, etc.Fourthly, I carry out the binary image with the close of morphologic. Fifthly, the coordination of the target will be obtained. Then I make many patches between the areas which are in cross rectangles. Then I will relocate the target. If the two targets unite into one target, I will consider the two targets as one target. I call this method“making patch”. However, we must pay attention that processed image was compressed. Then the coordination must be twice of original coordination and get the targets in original images. This algorithm skillfully handles the original images with compression and combination. This algorithm greatly improves the efficiency of program and reduces the size of using memories. Due to this algorithm which takes the image from the original images, there are no influences in the course of compression。Meanwhile,I take double thresholds, so there are no influences between large targets and small targets.Secondly, in the respect of image pretreatment and extracted feature: In order to boost up the reliability and consistency of feature, I mainly extract feature by means of geometry shape、imitation ellipse and frequency analysis. Since there are great differences between large target and small target, I will distinguish them with area. Then we discuss them irrespectively:For one thing, we will discuss the feature of the small target. First of all, we must take binary images into the small image. Although the binary images are mentioned in the segmentation images which process the whole images, we will aim at the small images in this time. Due to much similarity in shape between crystals which are diamonds and erythrocytes which are circle in small targets, we must ensure the binary image edges not to be distorted. The algorithm operations distorted the binary image edges with much severity by amounts of experiments. So we take the binary image according to histogram which greatly keeps the binary image non-distorted. Then we remove the noises, fill the inner field including close and open with grey valve zero. Then we extract the feature such as the ratio of rotundity, square of rotation ,cross feature and frequency ratio of rotundity based on FFT, which are innovation of this dissertation.For another thing, we will discuss the feature of the large targets. First of all, we get the edge information of targets with PREWITT algorithm operation. Then we get binary images as method of small targets, process with close of morphologic, remove the noises, fill hole and so on. At last, we acquire the geometry features such as the ratio long-short axes of ellipse imitation the ratio long-short axes of rotation , the ratio of rotundity, the ratio of stretch and so on. Thirdly, in the respect of the design of class facility:The class facility of this dissertation introduces neural net system on bin-tree. At first, the similarity targets are cursorily assorted by means of bin-tree. Then neural net system is used to accurate identification. This algorithm improves veracity of class on neural net .The neural net system has much better ability in the way of learning by itself and learning a certain something by heart. However, there is a fatal defect which is that it requires to be trained with a very long time. How it finishes training during a short time turns into one of problems noticed by researchers. If the position of output function of neural cell in saturation field and non-saturation field is justified, the learning efficiency of the neural net must be greatly improved.In order to accomplish this algorithm, the simple method is that the slope of output function of the standard BP algorithm is transformed from invariableness to variableness. Meanwhile, the standard BP algorithm acts on revision of the slope. Namely, the bigger slope coefficient makes the position of saturate field forwards, or else the smaller slope coefficient makes the position of saturate field backward. Such is the BP algorithm of adapting coefficient to itself.Fourthly, in the respect of program:The program design method of the orient object on C++ is adopted. For this language has some better excellences such as maintainability, transplantability, agility, encapsulatability, modularity, integrity and so on.3. Conclusions.This dissersation has researched and accomplished the segmention and feature extratation of the microscopic urinary image, and the recognition of the neural net on bin-tree. During the course of the urinay microscopic image process, the efficency of algorithm and the ratio of accuracy and identification are surrounded. Finally, the quality of segmentation has been boosted up greatly.The efficency of segmentation has been improved obviously. The consistency, stability ,the distinguishing ability, anti-jamming and the efficency of feature algorithm have been improved greatly. The training time of the neural net is reduced greatly, the ratio of accuracy and the ratio of recognition of cell has been increased greatly.At last,the efficency of expectation has been reached.更多还原