基于同层多尺度核卷积神经网络的单细胞图像分类研究

发布时间：2019-07-09 11:47

【摘要】：使用机器学习方法进行图像的分类是当下热门的研究方向之一,Google的AlphaGo与人类的围棋大战,更是促进了机器学习的进一步发展,卷积神经网络是应用广泛的机器学习算法,当下在金融风险预警,企业决策,字符识别等领域都有较多的应用。医疗影像技术随着科技的进步在临床上得到广泛应用,使得很多疾病可以更早发现、更早治疗,医学图像分类主要依靠人眼观察,耗时、耗力、且容易受到医生主观因素影响,医学细胞图像的计算机分类也成为研究的热点和难点。本文研究对机器学习算法——卷积神经网络模型进行改进并应用于单细胞图像分类,完成的主要研究工作如下:1、阐述了卷积神经网络的基本理论,介绍了从感知机模型演变到卷积神经网络的过程,以及Softmax分类器等理论知识,为模型的建立和改进提供理论基础。2、针对公开的HEp-2细胞图像库和宫颈细胞图像库中图像数量较少,且大小尺寸不一,无法直接训练卷积神经网络模型;本文研究使用缝裁剪法、高阶插值法对图像进行缩放,使得单细胞图像具有相同的尺寸,同时使用裁剪、对比度、亮度变换、旋转、规范化等方式对于小数据集进行数据提升,扩大了训练集合。3、针对经典的卷积神经网络模型不能有效的进行单细胞图像分类的问题,设计了一种改进的卷积神经网络模型,通过批技术、局部响应归一化,Softmax等六个方面改进了卷积神经网络模型。4、为了进一步提高卷积神经网络模型对单细胞图像的分类识别能力,考虑到人类视觉过程中感受野并非固定不变,设计了一种改进的卷积神经网络模型——同层多尺度核卷积神经网络,通过增加多尺度感受野、使用ReLUs、变更核函数数量等方式对卷积神经网络模型进行了改进。5、使用Google的Tensorflow框架编写改进的卷积神经网络模型,并通过ICPR2012 HEp-2数据集和宫颈细胞集对改进模型进行分类仿真测试。仿真结果表明:本文设计的两个模型具有较好的鲁棒性和抗扰性,对于图像残缺、对比度亮度不同以及旋转等都仍然能够很好的完成HEp-2单细胞图像的六分类和宫颈细胞图像的二分类识别。
文内图片：

图片说明：图１－２三类正常宫颈细胞示例逡逑
[Abstract]:The classification of images by using the machine learning method is one of the most popular research directions. Google's AlphaGo and the human's go war have also promoted the further development of machine learning. There are many applications in the fields of enterprise decision-making and character recognition. The medical image technology is widely applied in clinic with the progress of science and technology, so that many diseases can be found earlier and earlier, and the medical image classification mainly relies on the human eye to observe, time-consuming and power-consuming, and is easy to be influenced by the subjective factors of the doctor, The computer classification of medical cell images has also become the hot and difficult point of the study. In this paper, the machine learning algorithm _ convolution neural network model is modified and applied to the single-cell image classification. The main research work is as follows:1. The basic theory of the convolution neural network is described, and the process of evolution from the perceptron model to the convolution neural network is introduced. and based on the theoretical knowledge of the Softmax classifier, the theoretical basis is provided for the establishment and improvement of the model, and the convolution neural network model cannot be directly trained for the disclosed HEp-2 cell image library and the cervical cell image library, and the convolution neural network model cannot be directly trained; and the joint cutting method is researched in the paper, the high-order interpolation method is used for scaling an image so that the single-cell images have the same size, In order to solve the problem of single-cell image classification in the classical convolution neural network model, an improved convolution neural network model is designed, and the convolution neural network model is improved by using six aspects such as batch technology, local response normalization, Softmax and the like. in ord to further improve that ability of the convolution neural network model to classify the single-cell image, considering that the receptive field is not fixed in the human visual process, an improved convolution neural network model _ co-layer multi-scale nuclear-convolution neural network is designed to increase the multi-scale receptive field, In this paper, the convolution neural network model is improved by using ReLUs and the number of kernel functions.5. The modified convolution neural network model is written by using the Google's Tensorflow framework, and the improved model is classified and simulated by the ICR2012 HEp-2 data set and the cervical cell set. The simulation results show that the two models designed in this paper have good robustness and smoothness, and can still well complete the six-classification of the HEp-2 single-cell image and the two-class identification of the cervical cell image for the incomplete image, the contrast brightness and the rotation.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】