Research On Image Classification Based On Few-Shot Learning

Few-shot learning aims to learn and generalize from the limited samples to distin-guish new conceptions.Nowadays,metric learning and meta-learning have shown sig-nificant potential in few-shot learning.Among them,the quality of the feature extractor plays the most important role in the few-shot classifier performance.However,classic supervised training and instance contrast learning both exist different degrees of issues.Supervised training relies on label-driven that destroys the rich fine-grained semantic in-formation of images.Instance contrast learning addresses this problem to some extent.However,the definition of positive and negative sample pairs in the pre-task of contrast learning is still limited to local attributes and lacks global class information.In our paper,we propose the label-mask matrix as a supervised signal.This method considers the cor-relation of features between different images of the same class on the basis of instances.Secondly,we design a background confusion method as pre-task to accumulate foreground knowledge that is more suitable for few-shot learning by suppressing background-related shortcut knowledge.Finally,to alleviate the unstable meta-learning optimization during the few-shot classification stage,we add a multi-task setting to the metric-based approach.Additional auxiliary tasks in the embedding space could provide double constraints on the model parameters.Extensive experiments demonstrate that our proposed method is not only effective in improving the benchmark dataset for mainstream few-shot image classi-fication but also in cross-domain classification tasks.To summarize,our research work makes the following contributions and innovations:1.To address the lack of global class information in instance contrast learning,we propose a label-mask matrix and background confusion algorithm,corresponding to supervised contrast learning and its pre-task,respectively.The key difference be-tween supervised contrast learning and instance contrast learning is that the positive samples contain not only background confusion samples of the target data point but also other samples of the same class.Our proposed method achieves the state of the art in the baseline on the mini-image Net dataset.2.To address the problem of instability in parameter optimization of few-shot clas-sifiers over meta-learning,we propose to add multi-task settings to the few-shot learning classifier,including predicting image rotation angles,predicting the rela-tive positions of image patches,and image embedding space clustering.The auxil-iary tasks can help the feature extractor to learn more general semantic information to a certain extent.In the few-shot classification stage,the cosine classifier can better compute the distances in the embedding space.Through extensive experi-mental comparisons,the image embedding space clustering task performs best in all benchmark datasets.3.Based on the above-proposed methods,we present the experience results of the model on mini-image Net,CIFAR-FS,CUB-200-2011,and cross-domain datasets in detail.At the same time,we quantitatively analyze the label-mask matrix,back-ground confusion module,and multitasking settings that contribute to the perfor-mance of the model.Finally,to help readers better understand,we visualize the feature space and background confusing modules of supervised contrast learning.