FAHM: a fast asymmetric hashing method for remote sensing image retrieval using a lightweight deep representation learning network

Deep asymmetric hashing has emerged as an effective technique for large-scale remotesensing image retrieval (RSIR), primarily by alleviating the feature extraction burden associated with large-scale image databases. However, a significant limitation of prevailing methods is their reliance on computationally intensive, dense convolutional operations, which leads to substantial computational overhead and protracted training cycles. To address these challenges, this paper introduces a Fast Asymmetric Hashing Method (FAHM) designed to improve feature extraction efficiency while simultaneously enhancing the representation of intrinsic image structures. The core of FAHM is a novel lightweight feature extraction network, termed FSNet, which leverages partial convolution (PConv) to reduce redundant computations and integrates a bespoke structural embedding module to effectively capture and encode salient geometric features within remote sensing imagery. Furthermore, we propose an innovative hash code learning strategy that jointly optimizes a classification loss and a similarity-constrained loss. This dual-supervision mechanism ensures the generation of hash codes that are not only semantically consistent but also highly discriminative. Comprehensive experiments on three public benchmark datasets demonstrate the superiority of the proposed method. FAHM not only achieves state-of-the-art retrieval accuracy but also demonstrates a marked improvement in retrieval efficiency, outperforming baseline methods by a significant margin.