基于BP神经网络的红细胞识别分类方法研究与系统实现
发布时间：2018-07-28 09:37
【摘要】：红细胞的形态特征分类与识别在医学研究上有着重要的意义。目前,许多医院和科研院所对血液红细胞的分类与识别都是采用显微镜下人工观测细胞,再对其形态进行统计。这种方法只能对红细胞的形态和数量等信息做大致的统计和记录,如果要精确记录红细胞的面积、形状、圆形率等信息,人工观测的方法很难做到精确而细致的记录,同时对医护和科研工作者是极大的精力消耗。如果有一套健全的红细胞识别与分类系统,能够对它进行全自动化处理包括精准的形态分析、图表信息统计等一系列处理,这将极大的提高红细胞的检测准确度和医疗技术人员的操作效率。本文从医院及科研院所的实际检测需求出发,设计并实现了基于BP神经网络的红细胞识别与分类系统。具体包括以下工作:1.使用双边滤波平滑图像,采用BM3D算法去除图像的高斯白噪声并通过对图像进行灰度形态学操作增强目标与背景的对比度,同时突出图像边缘。预处理和增强模块是为了提升图像品质,为接下来的分割做铺垫。2.在先后比较了阈值分割、Canny分割和分水岭分割等方法之后,根据对分割算法的大量无差别测试,并分析、比较分割结果,选择了一种基于标记的分水岭分割算法,这个算法既能有效地分割红细胞,同时也能保证边缘特征信息不被大量丢失。分别提取红细胞的周长、面积、圆形度、矩形度及傅里叶形状描述子等特征,这些都是具有一定区分度的形状特征描述。3.用BP神经网络对红细胞进行识别与分类,对红细胞形状数据进行归一化和交叉验证等操作。最后训练得出一个识别率达到93%以上的BP神经网络。4.根据算法完成的任务,将算法处理过程逐一划分为功能独立又相互协作的功能模块,同时保持了算法的可扩展性。系统采用三段式界面,这种界面组织方式灵活并且可定制化,非常符合软件开发要求。对系统的功能模块进行了编码实现与功能测试,对红细胞检测结果进行图表化显示。5.比较了BP神经网络和决策树在红细胞识别的准确率、计算复杂度和算法可扩展性的优劣,得出前者比后者优越的结论。
[Abstract]:The classification and recognition of the morphological characteristics of red blood cells are of great significance in medical research. At present, the classification and recognition of red blood cells in many hospitals and research institutes are observed artificially under microscope, and their morphology is counted. This method can only make general statistics and records of red blood cell shape and quantity. If we want to accurately record the area, shape and roundness of red blood cells, it is very difficult for manual observation methods to record accurately and meticulously. At the same time, the health care and research workers are a great energy consumption. If there is a sound red blood cell recognition and classification system, it can be processed with full automation, including accurate morphological analysis, chart information statistics, and a series of processing. This will greatly improve the detection accuracy of red blood cells and the operational efficiency of medical technicians. In this paper, the recognition and classification system of red blood cells based on BP neural network is designed and realized according to the actual testing requirements of hospitals and scientific research institutes. Include the following work: 1. Using bilateral filtering to smooth the image, BM3D algorithm is used to remove the Gao Si white noise of the image, and the contrast between the target and the background is enhanced by gray-scale morphological operation, and the edge of the image is highlighted at the same time. The preprocessing and enhancement module is designed to improve the image quality and pave the way for the next segmentation. After comparing threshold segmentation with Canny segmentation and watershed segmentation, a watershed segmentation algorithm based on marking is selected according to a large number of undifferentiated tests and analysis of segmentation results. This algorithm can not only effectively segment red blood cells, but also ensure that the edge feature information is not lost. The circumference, area, roundness, rectangle and Fourier shape descriptors of red blood cells were extracted respectively. BP neural network is used to identify and classify red blood cells, and to normalize and cross-verify the shape data of red blood cells. Finally, a BP neural network with a recognition rate of 93% or more was obtained. According to the tasks accomplished by the algorithm, the algorithm processing process is divided into functional independent and cooperative function modules one by one, while maintaining the scalability of the algorithm. The system adopts a three-segment interface, which is flexible and customizable, and meets the requirements of software development. The function module of the system is coded and tested, and the result of red blood cell detection is graphically displayed. 5. The accuracy, computational complexity and expansibility of BP neural network and decision tree in erythrocyte recognition are compared, and the conclusion that the former is superior to the latter is obtained.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R446.1;TP391.41
[Abstract]:The classification and recognition of the morphological characteristics of red blood cells are of great significance in medical research. At present, the classification and recognition of red blood cells in many hospitals and research institutes are observed artificially under microscope, and their morphology is counted. This method can only make general statistics and records of red blood cell shape and quantity. If we want to accurately record the area, shape and roundness of red blood cells, it is very difficult for manual observation methods to record accurately and meticulously. At the same time, the health care and research workers are a great energy consumption. If there is a sound red blood cell recognition and classification system, it can be processed with full automation, including accurate morphological analysis, chart information statistics, and a series of processing. This will greatly improve the detection accuracy of red blood cells and the operational efficiency of medical technicians. In this paper, the recognition and classification system of red blood cells based on BP neural network is designed and realized according to the actual testing requirements of hospitals and scientific research institutes. Include the following work: 1. Using bilateral filtering to smooth the image, BM3D algorithm is used to remove the Gao Si white noise of the image, and the contrast between the target and the background is enhanced by gray-scale morphological operation, and the edge of the image is highlighted at the same time. The preprocessing and enhancement module is designed to improve the image quality and pave the way for the next segmentation. After comparing threshold segmentation with Canny segmentation and watershed segmentation, a watershed segmentation algorithm based on marking is selected according to a large number of undifferentiated tests and analysis of segmentation results. This algorithm can not only effectively segment red blood cells, but also ensure that the edge feature information is not lost. The circumference, area, roundness, rectangle and Fourier shape descriptors of red blood cells were extracted respectively. BP neural network is used to identify and classify red blood cells, and to normalize and cross-verify the shape data of red blood cells. Finally, a BP neural network with a recognition rate of 93% or more was obtained. According to the tasks accomplished by the algorithm, the algorithm processing process is divided into functional independent and cooperative function modules one by one, while maintaining the scalability of the algorithm. The system adopts a three-segment interface, which is flexible and customizable, and meets the requirements of software development. The function module of the system is coded and tested, and the result of red blood cell detection is graphically displayed. 5. The accuracy, computational complexity and expansibility of BP neural network and decision tree in erythrocyte recognition are compared, and the conclusion that the former is superior to the latter is obtained.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R446.1;TP391.41
【参考文献】
相关期刊论文 前10条
1 汪祖辉;孙刘杰;邵雪;;一种改进的三维块匹配图像去噪算法[J];包装工程;2016年21期
2 路，
本文编号：2149677
本文链接：https://www.wllwen.com/linchuangyixuelunwen/2149677.html
最近更新
教材专著
