Transformer-Based Distillation Hash Learning for Image Retrieval

Abstract

In recent years, Transformer has become a very popular architecture in deep learning and has also achieved the same state-of-the-art performance as convolutional neural networks on multiple image recognition baselines. Transformer can obtain global perceptual fields through a self-attention mechanism and can enhance the weights of unique discriminable features for image retrieval tasks to improve the retrieval quality. However, Transformer is computationally intensive and finds it difficult to satisfy real-time requirements when used for retrieval tasks. In this paper, we propose a Transformer-based image hash learning framework and compress the constructed framework to perform efficient image retrieval using knowledge distillation. By combining the self-attention mechanism of the Transformer model, the image hash code is enabled to be global and unique. At the same time, this advantage is instilled into the efficient lightweight model by knowledge distillation, thus reducing the computational complexity and having the advantage of an attention mechanism in the Transformer. The experimental results on the MIRFlickr-25K dataset and NUS-WIDE dataset show that our approach can effectively improve the accuracy and efficiency of image retrieval.