Oracle Character Image Retrieval by Combining Deep Neural Networks and Clustering Technology

In this paper, we study the description and retrieval of oracle character images. A great challenge is that there is a big difference between the variants of the same oracle character in character patterns, bringing deep negative impact on the retrieval of oracle character images. To solve this problem, in this paper, an image retrieval method is proposed by combining deep neural networks (DNN) and clustering technology. Firstly, for making full use of the powerful image representation ability of DNN, the deep convolutional neural network is used to extract character-level image features. Then, in order to solve the problem of retrieving variants of oracle characters, the Kmeans++ clustering algorithm is used to divide the oracle character dataset into different subsets of variants. Finally, based on the newly divided data set, a fully connected neural network is trained to get the variant-level features, which are used for the retrieval of oracle character images. Compared to the traditional methods and the DNN-based methods, this method achieves the optimal results on the data set of oracle character images.