Frequency Decoupling Enhancement and Mamba Depth Extraction-Based Feature Fusion in Transformer Hashing Image Retrieval

Jiayi Chen, Shuli Cheng, Liejun Wang, Yongming Li, Qiang Zou

Abstract: Highlights•We propose an adaptive frequency decoupling module via discrete cosine transform.•We first introduce Mamba to propose a hybrid state space module.•We introduce the concept of proxy and propose a novel global similarity proxy loss.•The FMTH model achieved high retrieval accuracy of 98.31% and a 7.29% improvement.