Oracle Bone Inscriptions Image Retrieval Based on Metric Learning

The goal of oracle bone inscriptions image retrieval is to find the most similar image in the oracle database. This technology provides valuable tools and methods for scholars to promote the digital research and development of oracle bone inscriptions. However, there are many challenges in the practical application of oracle bone images, such as severe noise interference, sample loss, and low similarity within the same class, while the high similarity between the different classes. To address these issues, this article first introduces a denoising process for oracle bone images to reduce the impact of noise on image retrieval. Secondly, a multi-strategy data augmentation method is also adopted to expand the fewer representative samples and enhance their diversity. Finally, this article proposes a metric learning-based Oracle image retrieval method that compares query images with images in image databases. To improve the retrieval accuracy, we have designed a feature extraction network specifically tailored for oracle bone image retrieval. This network includes a module for multi-scale convolution calculations of inputs, as well as a spatial pooling component for generating oracle image retrieval vectors. In the two retrieval datasets utilized in this article, our proposed method achieved performance improvements of 4.92% and 4.59% respectively compared to before improvement. This advancement will contribute to the widespread application of oracle bone inscriptions image retrieval in academic research and other relevant fields.