Abstract: In classification-based pattern recognition tasks,the training process requires handling a large number of category samples.In practice,these samples exhibit a significant long-tailed distribution characteristic,posing substantial challenges for such tasks.The challenges brought by long-tailed distribution mainly manifest in two aspects:imbalanced feature space,and difficulty in focusing on hard samples in the tail regions.To address these issues,a category-aware semi-supervised knowledge distillation model is proposed,which comprises two core components:balanced semi-supervised knowledge distillation and balanced category-aware learning.The former employs semi-supervised knowledge distillation to achieve a more balanced feature space.The latter integrates a category-aware margin loss function with a delayed hard sample learning activation loss function,improving classifier performance and enhancing focus on hard samples.All experiments were conducted on five benchmark datasets,including CIFAR10-LT,CIFAR-100-LT,ImageNet-LT,iNaturalist2018,and Places-LT.Notably,on ImageNet-LT,the proposed model achieved a Top-1 accuracy of 57.5%,outperforming other models.

Key words: semi-supervised, category-aware, long-tailed classification