Improved coordinate attention and prototype rectification for few-shot learning
In few-shot scenarios,accomplishing the task of identifying new categories with minimal labeled data is highly challenging.Due to data limitations,few-shot learning methods based on traditional prototype networks exhibit significant biases in the generated class prototypes.Currently,most research relies on convolutional neural networks as feature extractors,capturing only local relationships and failing to comprehensively extract sample information.To tackle those issues,a novel few-shot learning approach is proposed,integrating improved coordinate attention mechanisms with embedding propagation and prototype rectification.The model enhances the spatial and channel information linkage of images in horizontal and vertical directions through improved coordinate attention mechanisms,obtaining more accurate image features from samples.Employing the embedding propagation algorithm smoothes embedded features,predicts query set labels via label propagation,and selects query set features to rectify support set prototypes based on Euclidean distance weighting,resulting in more representative class prototypes.Experimental validation on the miniImageNet,tieredImageNet,and CUB datasets demonstrates the method's favorable outcomes in comparison to other techniques.These findings suggest promising prospects for addressing few-shot learning challenges.
万方数据
维普
