A Study Of Semi-Supervised Methods Based On Feature Fusion Contrast Learning And Dynamic Confidence Intervals

In machine learning,supervised learning requires a large amount of labeled data to train a model,but in reality,labeled data is often difficult or expensive to obtain,so SemiSupervised Learning(SSL)is needed to improve the generalization ability of the model using unlabeled data.Among today’s semi-supervised methods,consistent canonical methods have achieved good results in the semi-supervised domain,which typically use unlabeled data to construct different input distributions and let the model make predictions under these distributions.The model is then encouraged to be consistent by minimizing the variability of the output under different distributions.Although consistent regular methods can effectively improve the generalization ability of the model,they do not sufficiently consider the structural information of the unlabeled samples themselves.To address this problem,two types of semisupervised learning algorithms based on consistency regularity are proposed in this paper,which work as follows:(1)semi-supervised method based on feature fusion contrast learning(FFCL-SSL).The existing semi-supervised learning methods in which are mainly based on the consistency regularity approach use unlabeled data to construct different input distributions and let the model make predictions under these distributions.However,these methods do not sufficiently consider the semantic information between unlabeled samples,i.e.,samples of the same kind should be more closely related to each other.Therefore,in this paper we propose a semisupervised model based on contrast learning,where the model encourages structurally similar data points to have higher semantic similarity.In this model,we use data augmentation to expand the unlabeled dataset,and feature blending of two randomly selected images to construct new images.Then,we use a dual-model structure for feature extraction from the dataset and weighted fusion of the output features of the model.Finally,we calculate the weighted fused feature vectors in comparison with the feature vectors of the blended images after model prediction.To verify the effectiveness of the present model,we conducted experiments on a benchmark dataset.The results show that the present model can obtain better classification performance compared with other semi-supervised learning methods.(2)Dynamic confidence interval-based semi-supervised method(DCI-SSL)Most of the traditional semi-supervised learning methods determine pseudo-labels by confidence level,which does not change,so incorrectly labeled instances may be selected when the confidence level is low,and correct instance labels will be lost when the confidence level is high.Therefore,in this paper,we introduce a dynamic filtering of pseudo-labels to improve the classification accuracy of the model.In this paper,we determine a confidence selector that calculates the latest confidence level at each iteration and uses the current confidence level to filter the pseudo-labels.The pseudolabels identified in this round are then computed with the loss of the feature vector output by the model.In this study,we select the CIFAR-10 dataset,which is widely used for image classification,as the experimental test set to verify the effectiveness of the semi-supervised model proposed in this paper.The experimental results show that our proposed model outperforms other methods in terms of accuracy and error rate.