A Multi-granularity Decision Fusion Method Based on Category Hierarchy

Jian-Xun Mi, Ke-Yang Huang, Nuo Li

Abstract: The increasing complexity of convolutional neural networks (CNNs) has greatly improved the accuracy of image classification. However, there are unbalanced visual similarities between image categories, which results in different difficulties in classification. The existing solutions construct hierarchical structures according to image labels and use hierarchical information of categories to improve the performance of the network by constructing tree classifiers, but there is a problem of error transmission. In this paper, a multi-granularity decision fusion method based on category hierarchy is proposed. The two-level hierarchical structure of the image label is constructed according to the traditional algorithm, and the accuracy of image classification is improved through the decision fusion method of coarse and fine granularity, which effectively avoids the inter-level transmission of error. Real public data sets are used for verification. Experimental results show that, compared with the original convolutional neural network, the proposed method achieves higher classification results on CIFAR-10, CIFAR-100, and SVHM data sets.