Research On Few-Shot Image Classification Based On Attention And Clustering

In recent years,deep learning has been widely concerned for its excellent perfor-mance.The deep learning of supervised learning mode requires a large number of high-quality labeled data training models.However,in real life,collecting a large number of labeled data will consume a lot of material and financial resources.At the same time,some samples can not be obtained in large quantities due to privacy problems,which limits the further development of deep learning.In order to overcome the re-quirements of deep learning model for a large number of training data and make the training model generalize quickly,scholars have proposed a Few-Shot Learning tech-nology,which aims to use a small amount of labeled data to train the model,so that the model can quickly adapt to new tasks.This thesis focuses on the task of Few-Shot Im-age Classification,mainly including Few-Shot Fine-Grained Image Classification and Few-Shot Semi-Supervised Image Classification.Firstly,the current task of Few-Shot Image Classification is mainly studied on coarse-grained datasets.Compared with coarse-grained datasets,the difficulty of clas-sifying fine-grained datasets lies in the large intra-class differences and small inter-class differences of fine-grained data sets.Therefore,the task of Few-Shot Fine-grained Im-age Classification is a more challenging task.Therefore,the covariance attention mech-anism is introduced to capture the unique characteristics of fine-grained images by calcu-lating the depth covariance attention between samples,so that the extracted fine-grained image features are more discriminative; At the same time,we also introduce a projection module,which uses the projection function to project the support set sample features into the query set space by calculating the projection function of the query set sample space,so as to find the task-aware support set features,so that the covariance attention module can pay attention to the task-aware features and enhance the generalization ability of the model.The effectiveness of the method is proved on four mainstream datasets.Secondly,we also research Few-Shot Semi-Supervised Image Classification.The unlabeled data has no constraint on the label space,that is,the unlabeled data may come from a new class that does not intersect the marked class.To solve this problem,we propose a top-down clustering(TDC)method.The proposed TDC explores the discrim-inative information of the unlabeled data via iteratively grouping the unlabeled features into coarse-to-fine clusters.We first warm up a feature extractor with both the labeled data and unlabeled data.Based on the features of the unlabeled data derived from the trained feature extractor,a cluster miner is applied to group these unlabeled features into some clusters to obtain their corresponding pseudo-labels.The samples close to the class centroids are annotated as confident instances and are merged into the labeled data to update the base model for the next iteration.Progressive clustering is implemented by enlarging the number of clusters during the multiple iterations of clustering.The effectiveness of the method is proved on three mainstream datasets.