Hierarchical Classification Boost Using Confidence Belief Propagation
Lingzhu Deng, Yunfeng Sui, Long Chen, Shixuan Zhao, Weiqian Liu, Zhi Cheng · 2020
Fine-grained classification (FGC) is a constant tough task due to high similarity between sub-classes. Most research ignores the fact that FGC problem is also a hierarchical classification problem. In many situations, it is not able to make classification decision for leaf nodes in hierarchical tree, due to the lack of information. Our research treats FGC problem as multi-label classification problem. In addition, a confidence belief propagation method is proposed to solve the inconsistence of multi-labels in hierarchical relationship. Experiments on bird classification dataset with three hierarchical levels show that, the proposed approach is able to improve hierarchical classification accuracy, and it is also able to provide accurate confidence score for identifying the indistinguishable cases.