Research On Multi-index Combined Active Learning For Image Classification

The progress of deep learning applications is partly due to the large amount of annotated data.However,it is not easy to obtain a large amount of labeled data in many application scenarios,and active learning is one of the effective methods to alleviate this problem.Active learning methods based on pre-designed selection strategies can obtain higher-performance models with less labeled data,but a single selection strategy cannot always maintain high efficiency throughout the learning stage.It is more reasonable to use complementary multi-index sampling strategies.This paper combines the classic uncertainty sampling strategy and the source domain difference sampling strategy to construct a dynamic combination strategy,and proposes a new active learning method CDU(active learning method based on Combining source-domain Difference and target-domain Uncertainty,referred to as CDU).The source domain differential sampling strategy,which does not require the source task sample data,but only the source task model and the source task sample feature Gaussian mixture model,can solve the problem of low learning efficiency of the target task model in the initial stage of transfer learning(in real application of transfer learning,it is usually easy to obtain the source task model,but difficult to obtain the source task data for comparison with the target task samples,which makes it possible to repeatedly label and learn samples that is similar or even the same as the source task samples during the process of fine-tuning the target task model,resulting in waste of labeling costs and low learning efficiency).The CDU method first uses source task model and source task sample feature Gaussian mixture model to calculate the difference between target task samples and source task samples(source domain difference);then the source task model is used as the target task initial model to start active learning iteration;In the process of active learning iteration,the scores based on the dynamic integration of source domain differences and target domain uncertainty are used to screen out the most valuable target task samples for labeling and learning of the model.In addition,in view of the problem of unbalanced samples among categories when using actively learning to select samples in multi-category classification,the CDU method introduces high-score-and-random hybrid sampling to select samples.To prove the validity and rationality of the CDU method,experiments were carried out on the cross-domain transfer data sets of cat and dog recognition and handwritten digit recognition.The experimental results show that,compared with baseline methods,the CDU method can improve the learning efficiency of the target model in the initial stage of transfer learning,make the model more efficient in the entire learning stage,and effectively reduce the cost of labeling.