Embedded Separate Deep Localization Feature Information Vision Transformer for Hash Image Retrieval

Abstract

The development of multimedia technology has led to an increasing number of images, making the search for similar images an urgent need in daily life. Hash image retrieval has gradually dominated the field of image retrieval due to its advantages of computational efficiency and high accuracy. Currently, image retrieval algorithms based on Convolutional Neural Network (CNN) and Vision Transformer (ViT) remain inadequate in extracting target category features, ignoring local fine-grained features, which affects retrieval accuracy. This paper proposes an Embedded Separate Deep Localization Feature Information Vision Transformer for Hash Image Retrieval. Firstly, based on the diversity of image feature scales, a Channel Separation Attention Embedding Block (CSAE Block) is designed within the deep semantic feature layer. This block not only extracts global features but also incorporates a separated local attention branch to capture local features of objects at various scales. This enhances the output of deep features, providing rich semantic information for discrete mapping. Secondly, we design a quantization function that promotes the discreteness of hash codes, forcing the discrete values of the model output towards ±1. This ensures that the binary output of the hash code is more stable and representative. Finally, we conduct extensive experiments with the proposed algorithm on four public image retrieval datasets: MS-COCO, NUS-WIDE, ImageNet and CIFAR-10, achieving excellent retrieval performance with accuracies of 93.94%, 89.26%, 92.54% and 96.63%, respectively.