A dustbin category based feedback incremental learning strategy for hierarchical image classification

Abstract

Hierarchical classification (HC) is an effective method to solve the problem of multi-class classification especially when the categories are organized hierarchically. However, HC models perform worse than flat classification (FC) models due to blocking, i.e. errors occur at the higher level will be propagated to the lower level. The first step in solving the blocking problem is to capture the blocked samples. In this paper, we present a novel HC model to capture the blocked samples by adding a virtual dustbin category for each middle-layer classifier, referred to as DBHC model. Furthermore, in order to improve the classification accuracy and accelerate the convergence process, we propose a feedback incremental learning (FIL) strategy to take into account the weights of samples, which can adjust the composition of training samples according to the test results of the previous training steps. Experiments on fashion image classification shows the superiority of the proposed model compared with several prior methods.