Patent image retrieval using transformer-based deep metric learning

Abstract

Intellectual property work covers a wide range of areas. In particular, prior art literature searching in the patent field requires finding documents that can be used to determine novelty and inventive steps from a vast amount of past literature. Concerning this search practice, research and development of a drawing search technology that directly searches drawings, and essential information about inventions, has long been desired. However, patent drawings are usually described as black-and-white abstract drawings, except in some countries, and their modal characteristics are very different from those of natural images, so they have yet to be explored. This study achieved higher accuracy than the previous ones by introducing InfoNCE and ArcFace in the DeepPatent (Kucer et al., 2022) dataset instead of the conventional Triplet. In addition, we developed an application that enables users to search for patent drawings using any images. Our architecture can be applied to patent drawings and many other modal-like drawings, such as mechanical drawings, design patents, trademarks, diagrams, and sketches.