Research Of Cervical Cell Lesion Image Classification And Recognition Based On Convolutional Neural Network

Currently,the traditional cervical cell recognition is mainly through the cell image segmentation,artificial design operator to select features,then identified by the classifier,in cervical cell segmentation and feature extraction,the use of such methods need to master certain pathological knowledge,sometimes,since the characteristics of artificial selection,features selection are not considered representative,leads to the poor performance of the recognition.in this paper,a convolutional neural network based on deep learning framework is applied to cervical image recognition.Convolutional neural networks are a new type of artificial neural network,combines artificial neural network with deep learning,and which are able to combines feature extraction and classification process,and that can find the local characters of data by its dominant traits,receptive field,weights sharing and sampling strategy,therefore,it has been applied to many image recognition tasks.The convolutional neural networks model are applied to image recognition task in cervical cells,compared with the traditional artificial feature extraction,the image can be input directly and feature is extracted independently in the paper,that can improve the intelligence level and efficiency of cervical cell image recognition.The main researches are completed as follows:1.In this paper,the theory,characteristics and structure of convolutional neural network were elaborated,to provide theoretical basis for model improvement.The neural network is improved based on LeNet-5 network model in this paper.Several network models with different style of layer connection by changed the parameters of the network convolution kernel were constructed,and applied them on cervical cell image recognition,compared the classification results of each model by the simulation experiment,the influence of different filters on the network performance is analyzed2.Explored the influence factors of network performance basis of above studies,modifying the convolutional kernel size,the pool of selection,the activation function,and enhance the database in order to carry out the contrast simulation experiment.The results show that the reasonable parameters and the choice of methods will improve the classification and recognition performance of the network,especially the increase of image databases to improve the performance of the network significantly.3.After analyzing the influence factors of the classification performance of convolutional neural network,summarized the law of the reasonable choice of parameters and methods,constructed an optimal network structure for cervical cell classification and recognition.A network which increase the number of filters is constructed in the paper,and add BN algorithm to the network as the BN layer since the BN algorithm would promote the speed of the network training and convergence,then the dropout method is used to suppress the neuron in the network,and the softmax is used as the classifier at last,and then classification and identification of cervical cells.The simulation results show that:An improved convolution neural network is used to achieve the two classification accuracy of 98.36%in this paper,and the recognition effect is better than ANN method,SVM method,KNN method,Bayesian method and linear discriminant method,the recognition rate is 12.21%higher than that of the traditional Bias method,which is higher than that of the artificial neural network(ANN)method,which has a certain practical value.