On Expert-Defined Versus Learned Hierarchies for Image Classification

摘要

For classification task involving a large number of classes, a decrease in recognition accuracy is observed for visually similar classes. We believe that forcing the model to learn appropriate features separately for each set of similar classes could improve classification performance. To justify our idea, we tried to improve classification performance by employing class hierarchy, which reflects visual similarities in data. More specifically, we used and compared two kinds of hierarchies to enhance classification performance of the model: (i) a hierarchy defined by experts (H-E), and (ii) a hierarchy created from performance results of a flat classifier and using DBScan clustering method (H-C). Moreover, we created a classification model that efficiently utilizes these hierarchies to learn appropriate features at different levels of the hierarchy. We evaluated the performance of the model on CIFAR-100 benchmark. Our results demonstrate that the hierarchical classification under H-C outperforms both H-E and the flat classifier.