基于卷积神经网络的体液细胞图像有形成分的特征识别方法研究

发布时间：2018-03-17 06:20

本文选题：体液细胞　切入点：特征提取　出处：《湘潭大学》2017年硕士论文　论文类型：学位论文

【摘要】：医学检验中,尿检是极为重要的一项检查项目。针对尿检项目的繁杂,如何利用现代计算机技术高效地识别尿液标本的有形成分,辅助检验医师进行诊断是一项长久的研究课题。要想利用计算机准确地识别出目标成分,通常来说有几个关键性的技术:电子显微镜自动聚焦标本,去除图像噪声,计算机分割目标,特征提取和识别。现实生活中已有利用数字图像技术来实现上述技术,其中特征的提取和识别总的来说就是从形状和纹理的角度,综合运用如不变矩法、傅里叶变换法、图割拓扑法以及BP神经网络法,都取得了一定的成效,也都有各自的一些不足。深度学习是现今比较热的一个研究方向,有浅层学习所不具备的一些优势,其中卷积神经网络是比较成熟的一种深度学习网络,在各领域都有一定的研究应用。论文基于卷积神经网络的原理,在AVE镜检实验平台上,针对体液细胞的有形成分,开展识别研究工作。论文首先介绍开题的由来,接着从特征提取、深度学习、卷积神经网络三个方面总结归纳了目前的研究现状。特征提取通常的做法是从图像形态外观上获得描述性的语义模型,再由语义模型转化建立数学模型,数学模型要对样本有一定的区分度,多个特征模型之间不能相互干扰,通过训练和测试看能否达到令人满意的识别度。本论文主要完成的工作是:(1)介绍了人工神经网络的原理和结构,从人工神经网络的基本结构上入手比较了与其他特征识别方法的优势所在,并以BP神经网络为典型,写明了BP神经网络是如何进行图像的特征识别,改进了基于BP神经网络的体液有形成分识别方法,分析了其识别优劣势。(2)卷积神经网络作为本论文的重点研究部分,其结构与浅层人工神经网络是存在区别的,训练方法也是难点所在。本文改进了LeNet-5网络模型,自制了适合卷积训练的训练库和测试库,求出网络各层的灵敏度,再用误差去一步步的调节网络权值和阈值,最终得到了红、白细胞的识别误差曲线。通过比较,发现卷积神经网络在特征提取上存在一定的优势。
[Abstract]:In medical testing, urine examination is an extremely important item. In view of the complexity of urine examination, how to use modern computer technology to identify the visible components of urine samples efficiently, Assistant examiners' diagnosis is a long-lasting research project. To accurately identify target components by computer, there are usually several key techniques: electron microscopy autofocus specimens, remove image noise, Computer segmentation, feature extraction and recognition. In real life, digital image technology has been used to achieve the above technology, in which feature extraction and recognition is generally from the perspective of shape and texture, the integrated use of such as moment invariant method, Fourier transform method, graph cut topology method and BP neural network method have achieved certain results, and have their own shortcomings. Deep learning is a hot research direction, and has some advantages that shallow learning does not have. Convolution neural network is a kind of mature deep learning network, which has certain research and application in various fields. Based on the principle of convolution neural network, this paper aims at the tangible components of humoral cells on the AVE microscope test platform. First of all, the paper introduces the origin of the topic, and then from the feature extraction, in-depth learning, This paper summarizes the current research situation in three aspects of convolution neural network. The usual method of feature extraction is to obtain the descriptive semantic model from the appearance of the image, and then transform the semantic model to establish the mathematical model. The mathematical model should have a certain degree of discrimination to the sample, and there can be no interference between the multiple characteristic models. The main work of this paper is to introduce the principle and structure of artificial neural network. Based on the basic structure of artificial neural network, this paper compares the advantages with other feature recognition methods, and takes BP neural network as a typical example, describes how BP neural network carries out image feature recognition. This paper improves the recognition method of physical component of body fluid based on BP neural network, and analyzes its advantages and disadvantages. The convolutional neural network is the key research part of this paper. The structure of the neural network is different from that of shallow artificial neural network. The training method is also difficult. This paper improves the LeNet-5 network model, makes the training library and test library suitable for convolution training, calculates the sensitivity of each layer of the network, adjusts the network weight and threshold step by step with the error, and finally obtains the red. Through comparison, it is found that convolution neural network has some advantages in feature extraction.
【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】：TP391.41;TP183

【参考文献】