Research On Long-tail Image Classification Based On Ensemble Model

The rise of AI related technologies promotes the development of many industries in human society.Computer vision is a representative task in the field of AI,and widely used in medical,industrial,military and other fields.Image classification is the basic task of computer vision.The research and practice of image classification is conducive to the vigorous development of computer vision task.The image classification task usually assumes that the classes in training dataset follow uniform distribution,this is an important reason why neural networks perform well in this task.However,in many real scenes,the classes in training dataset usually present a long-tailed distribution:the number of samples in head classes is rich and the number of samples in tail classes is rare.The image classification at this time is called long-tail image classification.In long-tailed image classification task,due to the gradient space dominated by head classes,the generalization of neural networks for tail classes decreases dramatically,the classification accuracy reduces significantly.Many previous works show that enhancing the diversity of ensemble models can effectively boost their generalization ability,but the mechanism behind this phenomenon remains to further analyse.Therefore,in the multi-classification scenario,this paper first theoretically extended analyses the reason why the generalization ability of ensemble models is superior to individual models,then proposes two solutions for long tail image classification,namely,Multiple Classification Student Model based on knowledge distillation and Multiple Contrastive Expert Model based on contrastive learning.Multiple Classification Student Model is an ensemble model,in which each individual model is good at identifying different classes,and uses the teacher-student scheme to learn from the knowledge of the teacher model to improve the generalization ability.Multiple Contrastive Expert Model is also an ensemble model.which can simultaneously consider the diversity and representation ability of the ensemble model.In terms of the diversity of the ensemble model,Multiple Contrastive Expert Model combines individual models with unique characteristics,thereby more accurately approximating the true class distribution;In terms of representation ability,this paper uses supervised contrastive loss to guide the feature training of the model,resulting in the feature space corresponding to the model showing intra class attraction and inter class exclusion.This enhances the representation ability of the ensemble model while simplifying the training of linear classifiers.Extensive experiments on four long-tailed image classification datasets demonstrate the proposed approaches achieve excellent performance.