Hierarchical Few-Shot Learning Based On The Class Hierarchical Granulation

The key issue to be solved in few-shot learning(FSL)is effectively using limited sample resources to build a more accurate classification model for the current classification and identification needs.FSL has attracted lots of attention due to its huge potential value created by its efficient learning ability in artificial intelligence.Although existing inter-class metrics approaches have made great progress,most existing metric learning methods mainly focus on metrics between fine-grained levels,relying on feature fusion between network layers to improve the feature distinguishability of samples.Since the scarcity of samples,the performance of sample features generated by single granular information or inter-layer fusion method needs further improvement.In this direction,we focus on category-level granulation to further mine the representation of granular information.Furthermore,through feature fusion,we study how to effectively exploit the relationship between the knowledge contained in the data and the granularity of fusion to improve classification performance.Specifically,it mainly includes the following three aspects of research:(1)Hierarchical few-shot learning based on data-driven hierarchical granulationMany existing few-shot learning assume that the classes are independent of each other and ignore their granularities relationship when building a classification model.To this end,we propose a data-driven strategy to construct a coarse-and fine-grained relation network,which leverages the spectral clustering method to cluster similar sample features and mine the relationship among different classes.(2)Hierarchical few-shot learning based on data and bidirectional knowledge driveThe main limitation of existing few-shot learning with feature fusion is that it only utilizes the inter-layer network for feature fusion,which ignores the potential feature representation in the intra-layer network.To this end,we propose a hierarchical metric FSL model with comprehensive feature fusion driven by data and knowledge based on intra-layer channel features and hierarchical class structure perspectives.(3)Hierarchical few-shot learning based on bidirectional driven cross fusionSince the scarcity of few-sample data brings great challenges,we propose a fine-grained method for bidirectionally-driven cross-fusion of data and knowledge.We focus on fine-grained modal fusion to effectively utilize the complementarity and consistency between granularities.We break the intra-class modal monotonicity to make the intra-class feature spatial distribution more compact and the inter-class feature spatial distribution more distinct.The proposed methods verify the inherent consistency,complementarity,and fusion of granular features in mining data.An effective granular hierarchical category structure is an effective means to improve small sample classification.