Research On Segmentation And Feature Extraction Algorithms Of Cervical Cell Image

Cervical cancer is a gynecological malignant tumor disease with a high incidence that threatens women's health.The research and analysis of cervical cells is of great significance in the diagnosis of cervical cancer.The accurate detection and classification of microscopic cell images can directly assist clinicians in reading.With the development of automation,information and intelligence,and the rapid development of the economy,people are paying more and more attention to physical health issues,and they also hope to quickly diagnose the existing problems.However,due to the complex and changeable images of cervical exfoliated cells,existing deep-learning-based cervical cancer cell recognition has problems such as low detection recognition rate and time-consuming inspection.Therefore,optimizing the structure of cervical cancer cell recognition network is a challenging task and research value,and it is also an important issue related to women's health.Traditional cervical cancer recognition algorithms are mainly based on support vector machines and cascaded multi-classifier fusion.These algorithms are relatively complex during the practical application,and will be affected by other external factors to make the detection results inconsistent with the requirements of the real scene.Nowadays,recognition technology based on deep learning has become a hot research topic and the core of machine learning field.The main research direction of this topic is the target detection algorithm based on deep learning.The basic model selected through analysis is SSD model,which has relatively high accuracy and detection speed.Based on this model,further improvement is made to achieve the expected effect of cervical cancer identification.The main contents and innovations of this article are as follows:1.For the problem that no public cervical cancer cell data set is provided for network training,the first job is to collect microscopic images of cervical exfoliated cells,and then according to the definition of cancerous cervical epithelial cells and normal cervical epithelial cells in medicine,use label Img The software marks the cervical exfoliated cells,and stores the position information and species information of the target cells in the format of.xml file.2.Based on the SSD model,the shallow network feature information is used to predict the target objects.During this period,it will face other problems due to the lack of deep semantic features.Due to the FPN(Feature Pyramid Networks)model of the Bottom-top and top-Bottom two pathways of Feature fusion model inspired by designed connection module,and through a large number of experiments based on three characteristics between convolution model-top to Bottom and top-Bottom features fusion,this particular convolution Conv7,Conv4?3 and Conv8?2 respectively,through the combination of Bottom-up and top down method to obtain strong semantic characteristics,Improved performance of target detection on the dataset.Experimental results showed that the improved SSD model improved the detection accuracy of cervical cancer cell recognition by 2%.3.For overlap cervical epithelial cells easy to leak this problem,based on the SSD model to join the network structure,on the basis of FPN will replace the largest soft?nms algorithm suppression algorithm,through the soft?nms a decay function to reduce the degree of confidence,if only a small part of the cell area,is it the original confidence will not be affected by too much,for the current object detection algorithm in multiple overlapping soft?nms have significantly increased the average accuracy,so that you can avoid test results appear false positives,avoid the occurring of the testers.Finally,the performance advantages of the improved model were compared with the other three algorithms in the experimental way.