Long-tailed Classification Method Based On Hierarchical Knowledge-driven

With the advent of big Internet data and artificial intelligence,deep convolution neural networks have succeeded in image classification,target detection,speech recognition,and other fields.However,the data classes in the real world present a long-tailed distribution,which usually shows that a small number of head classes have large training samples.In contrast,a large number of tail classes have limited training samples.The unbalanced distribution of classes makes the network model easy to deviate from the head class,resulting in poor classification performance of the tail class.Therefore,the classification accuracy of convolution neural networks in the distribution of long-tailed data cannot meet the practical application requirements.Improving the classification performance of convolution neural networks under the long tail data distribution has become an important research direction.The main contents are as follows:(1)Long-tailed classification with semantic hierarchical knowledge transfer: Aiming at the traditional convolution neural network training each target class independently,a multi-task convolution neural network based on semantic hierarchical knowledge transfer is proposed,which uses super-class knowledge to promote long-tailed learning.First,we construct the semantic super-class hierarchy based on the tail class to pay more attention to the tail class.Second,the multi-task convolution neural network trains super-class and target tasks to extract more extensive knowledge.Finally,a knowledge transfer strategy from super-classes to targets is designed to adaptively adjust the task weights and improve the classification accuracy of target classes.(2)Long-tailed classification with semantic and clustering hierarchical knowledge complementation: Aiming at the fact that the semantic hierarchy fails to fully explore the class relationship and transfer redundant knowledge to the tail classes,a multi-task convolution neural network based on semantic and clustering hierarchical knowledge complementation is proposed to transfer useful knowledge to the tail classes.First,semantic and clustering superclasses are integrated into the convolutional neural network as hierarchical knowledge to guide feature learning.Then,a knowledge complementation strategy is designed to jointly use these two kinds of super-class knowledge,where semantic knowledge acts as a prior dependency.Clustering knowledge reduces the negative information caused by excessive semantic dependency.In this way,the convolution neural network uses two complementary hierarchical structures and transfers useful knowledge to the tail classes improving the accuracy of the long-tailed classification.