Transformer-Based Unsupervised Deep Hashing Using Knowledge Distillation for Image Retrieval

With recent technological advancements, a vast amount of data is being generated online, making images an essential tool in almost every field today. In various domains such as industry, art, and science, images are used as key means of information transmission and analysis, highlighting the necessity and importance of image search. However, implementing technology that satisfies both speed and accuracy in image search remains a challenge, prompting research into various image hashing techniques. Traditional hashing methods aim to maximize efficiency by converting high-dimensional data into lower dimensions, but they have limitations in accurately reflecting complex image patterns. In this paper we propose a deep hashing model that combines a Transformer-based backbone with knowledge distillation and an optimized loss function. The proposed model demonstrates higher search accuracy than existing hashing methods across various benchmarks and proves its superior performance in image search through qualitative results.