The Research On Long-tailed Visual Recognition Algorithms

In the real-world,large-scale datasets often exhibit long-tailed distribution,i.e.,a few classes occupy most of the data,while most classes have rarely few samples.When the traditional visual recognition methods are applied to the long-tailed image data,there will be problems such as model failure,recognition accuracy plummeting and so on.By analyzing the characteristics and difficulties of long-tailed visual recognition problem,this paper proposed two innovative algorithms for modeling long-tailed distribution image data.A long-tailed recognition algorithm based on dual channel learning(DC-LTR)is proposed.In order to alleviate the problem of large variance of intra-class features caused by imbalanced learning methods,DC-LTR algorithm adds a few-shot learning channel based on metric learning in addition to an imbalanced learning channel.The imbalanced learning channel mitigates the model's preference for head classes,while the few-shot learning channel repairs the feature representation by pulling closer the samples from the same class and pushing away the samples from different classes.DC-LTR algorithm uses a curriculum learning strategy to dynamically adjust the importance of two learning channels during training,which effectively reduces the variance of intra-class features of tail data,improves the recognition ability of tail data and enhances the recognition performance of long-tailed data as a whole.A long-tailed recognition algorithm based on hierarchical learning(HL-LTR)is proposed.HL-LTR algorithm turns the problem of long-tailed recognition into a hierarchical super class learning problem,in which the lowest super class learning task is the original long-tailed recognition task,and the learning task is gradually becoming easier from the bottom level to the top level as the data distribution become gradually balanced.HL-LTR algorithm learns the task from the top level to the bottom level and uses the knowledge learned from the top super class to guide the learning of the bottom super class,which is mainly achieved by an attention mechanism guided learning module and knowledge distillation technology.The transfer of knowledge is mainly between the super class and its subclasses,which alleviates the problem of domain drift.Compared with the method of solving the most difficult task of long-tailed recognition directly,HL-LTR algorithm has better recognition ability for the head as well as tail data and achieves better performance by gradually learning from easy to difficult and by directed knowledge transfer.This paper compared the proposed DC-LTR algorithm,HL-LTR algorithm with several popular long-tailed recognition algorithms on multiple long-tailed image datasets.The experimental results show that the DC-LTR algorithm performs best on few-shot class which has a small number of samples,while the HL-LTR algorithm achieves state-of-the-art result on the whole dataset.