Research On Few-Shot Image Cllasification Methond Based On Semi-Supervisied Learning

Currently,deep learning has reached or even surpassed human recognition in image classification tasks in the field of computer vision.But deep learning requires a huge number of labeled samples,which not only limits the development in related fields that lack a large number of labeled samples,but also has a large gap compared with humans who can rapidly iterate and learn new classes.Therefore,research on few-shot image classification has emerged,which focus on how to guarantee the reliability of deep learning with a small number of labeled samples.Among them,the semi-supervised learning method can reduce the cost of data collection,which is also close to the human cognitive model.This study conducts a research on the few-shot image classification method based on semi-supervised learning.Today,the main idea of applying semi-supervised learning in the field of few-shot image classification is to give pseudo-labels to unlabeled samples before adding training,which brings the problem that the reliability of pseudo-labels cannot be determined and is difficult to measure.In the first research work of this paper,a Transformer-based joint spatial attention module is proposed using transductive semi-supervised learning.The query set is added to the training as unlabeled samples,and the spatial correlation matrix between each unlabeled and labeled sample will be found,which in turn calibrates the feature centers of the categories and avoids the problems caused by assigning pseudo-labels.In the second research work of this paper,a Transformer-based local and global feature fusion module is proposed,which focus on the the local and global features of the sample by inductive semi-supervised learning.A better feature extractor is finally learned and combined with the previous work to achieve better few-shot image classification results.Applying a small amount of labeled data and a large amount of unlabeled data,this study conducts extensive experiments on three commonly used few-shot image classification datasets and two kinds of implementation settings,and the proposed method achieves significant performance improvements.