基于卷积神经网络的宫颈细胞病变图像识别研究

发布时间：2018-01-20 00:41

本文关键词： 宫颈细胞病变识别卷积神经网络网络分类识别性能样本扩容 BN算法　出处：《广西师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：目前传统的宫颈细胞识别主要都是先经过细胞图像分割,人工设计算子选取特征,然后选用分类器进行识别。在宫颈细胞分割与特征提取阶段,使用此类方法需要掌握一定的病理医学常识,而且由于特征是人为选取,有时候选取的特征并不具有代表性,会导致识别效果不明显,因此本文将深度学习框架下的卷积神经网络应用于宫颈细胞识别的领域进行研究。卷积神经网络是将人工神经网络和深度学习相结合的一种新型人工神经网络,能够将特征提取与识别分类工作相结合,其最主要的特点是局部感受野、权值共享和空间子采样,能够提取数据的局部特征,因此在图像识别领域获得了广泛的应用。本文将卷积神经网络模型应用到宫颈细胞图像识别中,本文的方案具有图像可以直接输入,特征自主提取的特点,可以提高宫颈细胞图像识别的智能化水平与效率。本文完成的主要研究工作如下:(1)本文详细阐述了卷积神经网络的理论、特点和结构,为模型的改进提供理论基础。本文在LeNet-5模型的基础上,构造了若干个具有不同的层间连接方式的抽取特征的滤波器层的卷积神网络模型,并将这些模型应用到宫颈细胞图像的识别中,通过仿真实验比较各个模型的分类效果,分析了不同数量的过滤器对网络性能的影响。(2)在上文研究的基础上继续探究影响网络识别性能的因素,通过调整卷积神经网络的卷积核尺寸、下采样方法、激活函数以及扩增图像数据集来进行对比仿真实验。仿真结果表明,合理的参数及方法选择都会提高网络的分类识别性能,尤其是增加图像数据集对网络性能提升效果明显。(3)经过分析了卷积神经网络识别分类性能的影响因素之后,总结了合理选择参数以及方法的规律,构造了一个宫颈细胞图像分类识别性能最佳的网络结构。本文构造了一个增加卷积层过滤器数量的网络,并加入了 BN算法作为BN层,BN算法能够加快网络训练速度与网络收敛速度,然后加入dropout方法,随机抑制网络中的神经元,最后使用softmax作为分类器,对宫颈细胞进行病变分类识别。仿真实验结果表明:本文构建的改进卷积神经网络对宫颈细胞图像二分类识别率达到98.36%,识别效果优于ANN方法、SVM方法、KNN方法、贝叶斯方法和线性判别方法等多种方法,识别率比传统贝叶斯方法提高了 12.21%,比人工神经网络方法(ANN)提高了5.65%,具有一定的实用价值。
[Abstract]:At present, the traditional cervical cell recognition is mainly through cell image segmentation, artificial design operator to select features, and then select classifier for recognition. In the cervical cell segmentation and feature extraction stage. The use of this method requires a certain degree of common sense of pathology medicine, and because the feature is artificial selection, sometimes the selected features are not representative, which will lead to the recognition effect is not obvious. In this paper, the convolution neural network in the framework of deep learning is applied to the field of cervical cell recognition. Convolution neural network is a new type of artificial neural network which combines artificial neural network and depth learning. Feature extraction and classification can be combined, the most important characteristics of the local receptive field, weight sharing and spatial subsampling, can extract the local features of the data. Therefore, it has been widely used in the field of image recognition. In this paper, the convolution neural network model is applied to cervical cell image recognition. It can improve the intelligent level and efficiency of cervical cell image recognition. The main research work accomplished in this paper is as follows: 1) the theory, characteristics and structure of convolution neural network are described in detail in this paper. On the basis of LeNet-5 model, this paper constructs several convolutional network models of filter layer with different interlayer connection characteristics. These models are applied to the recognition of cervical cell images, and the classification effects of each model are compared by simulation experiments. This paper analyzes the influence of different number of filters on network performance. (2) on the basis of the above research, we continue to explore the factors that affect the network identification performance, and adjust the convolution kernel size of the convolution neural network. The simulation results show that reasonable selection of parameters and methods can improve the classification and recognition performance of the network. In particular, adding image data sets to improve the network performance is obvious. After analyzing the factors affecting the classification performance of convolution neural networks, the reasonable selection of parameters and the rules of the method are summarized. A network structure with the best performance of classification and recognition of cervical cell images is constructed. A network to increase the number of convolutional layer filters is constructed and BN algorithm is added as the BN layer. BN algorithm can speed up the network training speed and network convergence speed, then add dropout method, randomly suppress the neurons in the network, and finally use softmax as classifier. The result of simulation experiment shows that the improved convolution neural network is better than ANN method in the recognition rate of cervical cell image. The recognition rate of SVM method is 12.21% higher than that of traditional Bayesian method, such as Bayesian method, Bayesian method and linear discriminant method. Compared with the artificial neural network (Ann) method, it is 5.65% higher than the artificial neural network (Ann) method, and has certain practical value.
【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】：R737.33;TP391.41;TP183

【参考文献】