Application Of Image Classification Method Based On Semi-supervision

Semi-Supervised Learning(SSL)can use a small amount of tag data for training,which solves the cost problem caused by tag data.Therefore,it has attracted many scholars’ attention in different fields.Especially in areas where it is difficult to obtain labeled data,this advantage is difficult to replace.Semi-Supervised Learning has many tasks and fields.This paper mainly discusses how Semi-Supervised Learning improves clustering effect and how Semi-Supervised Learning is applied to medical image classification tasks of dermatology.Through in-depth research and improvement of existing excellent algorithms,some of their problems are solved,and the performance of related semi supervised models is further improved.The main work and innovation points during the period are as follows:(1)Provided a HFC image classification method.Firstly,in response to the issue of RUC(Unsupervised Image Clustering With Robust Learning)not taking into account the impact of data similarity on its pseudo label sampling strategy,a method based on confidence and distance is proposed as the pseudo label sampling strategy,which combines the two strategies to improve the probability of filtering to the correct label.At the same time,the data augmentation method in FixMatch was used to replace the random augmentation in Mix Match in RUC,highlighting the combined effect of strong and weak augmentation.Finally,the Tri training method is used instead of Co training,which guides the third classifier through two classifiers for training,so that the model is no longer limited to the dual view dataset.The HFC method was tested on the CIFAR-10,CIFAR-100,and STL-10 datasets,verifying that compared to the RUC model,the HFC method improves accuracy,is more robust,and can improve the problem of overconfidence in cluster prediction.(2)A semi supervised classification method FKC based on adaptive threshold strategy and kernel norm regularization was proposed.In response to the problem of the semi supervised model FixMatch relying on a fixed threshold to filter unlabeled data without distinguishing the learning difficulty of different classes,FKC used a scaling threshold method to provide an adaptive threshold for the model,reducing the learning difficulty.Meanwhile,in response to the problem of excessive confidence in FixMatch,FKC adopts kernel norm regularization to increase the diversity of class distribution.Finally,a label smoothing operation was added to reduce the impact of noisy data on the model.FKC conducted experiments on four data sets,CIFAR-10,CIFAR-100 and STL-10,and achieved satisfactory results.Compared with FixMatch,FKC has a faster Rate of convergence and a more diverse class distribution in the late training period.(3)A prototype system for multi category medical image classification has been designed and implemented.The system includes several parts,including input data,data preprocessing,model training,and result analysis.The ISIC2018 skin lesion diagnosis dataset is used to train the model.After training,the model can predict the input images,which has certain feasibility and practical significance.