Research On Pseudo-label Method For Semi-supervised Image Classification

Semi-supervised learning(SSL)provides an effective method to improve classifier performance for unlabeled data,and various pseudolabel methods have shown excellent performance on classification task.However,this paper suggests that the existing fixed-threshold pseudo label strategy cannot provide a correction mechanism to alleviate the confirmation bias introduced by incorrect pseudo labels.In addition,the fixed-threshold design does not effectively utilize unlabeled data,which will lead to lower model performance and slower convergence.In consideration of the above concerns,this paper will improve the semisupervised classification performance from two points of view:pseudo label quality assessment and generation.First,this paper designs an adaptive thresholding strategy based on sample temporal prediction.The adaptive thresholding maintains the historical predictions of samples over a time series and calculates the base thresholds describing the learning level of the samples and the temporal uncertainty indicating the degree of model recognition.This dynamic threshold design enables the recall of valid hard samples and correction of noisy pseudo labels from two perspectives:reducing noisy pseudo labels and designing a correction mechanism.It also provides criteria for pseudo label quality assessment consistent with sample difficulty and model recognition ability.Then,to further improve the accuracy of pseudo label generation,this paper designs a multi-crops masking strategy to address the lack of the number of sample data augmented views of the current method,which encourages the model to perform diverse target feature learning in the early training phase.Combining the adaptive thresholding strategy and the multi-crops masking strategy,this paper proposes the TeST(TemporalStable Thresholding)method to improve the semi-supervised classification performance from pseudo label quality assessment and generation.Experiments show that the TeST method achieves more advanced classification performance on three mainstream semi-supervised learning classification benchmarks,especially when the labeled data are scarce.Extensive ablation experiments and visualization results demonstrate the performance improvement of each module of the TeST and its superiority on semi-difficult sample mining and model recognitive uncertainty reduction.