Scholarlink

<p>This paper presents a supervised hashing framework built on a Siamese architecture, where gated residual connections substantially enhance the quality of compressed representations. The model employs ConvNeXt-Base as the backbone, combining the inductive biases of convolutional networks with modern architectural principles inspired by transformers. In the hash-generation pathway, the learned representation must simultaneously preserve essential identity information and apply nonlinear transformations for effective compression and separation. To achieve this, a learnable gate is introduced between two complementary branches: (1) a shallow identity/residual branch that preserves the core feature structure extracted by the backbone, and (2) a deeper transformed branch that performs nonlinear projection and disentanglement. The adaptive gating mechanism dynamically balances these two paths, enabling the network to retain discriminative local and global cues while suppressing irrelevant variations. As a result, the proposed design produces stable, semantically consistent, and highly discriminative binary hash codes. The Siamese network is trained using Triplet Loss to enforce similarity preservation in Hamming space. Extensive experiments on fine-grained benchmarks (CUB-200-2011 and NABirds) as well as a coarse-grained dataset (CIFAR-10) demonstrate that the proposed framework consistently outperforms or matches state-of-the-art hashing methods, validating its robustness and generalization across different levels of semantic granularity.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing image retrieval via Siamese network-based hashing with gated residual connections

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Alireza Nazari,
  • Kambiz Rahbar,
  • Ziaeddin Beheshtifard,
  • Maryam Khademi

摘要

This paper presents a supervised hashing framework built on a Siamese architecture, where gated residual connections substantially enhance the quality of compressed representations. The model employs ConvNeXt-Base as the backbone, combining the inductive biases of convolutional networks with modern architectural principles inspired by transformers. In the hash-generation pathway, the learned representation must simultaneously preserve essential identity information and apply nonlinear transformations for effective compression and separation. To achieve this, a learnable gate is introduced between two complementary branches: (1) a shallow identity/residual branch that preserves the core feature structure extracted by the backbone, and (2) a deeper transformed branch that performs nonlinear projection and disentanglement. The adaptive gating mechanism dynamically balances these two paths, enabling the network to retain discriminative local and global cues while suppressing irrelevant variations. As a result, the proposed design produces stable, semantically consistent, and highly discriminative binary hash codes. The Siamese network is trained using Triplet Loss to enforce similarity preservation in Hamming space. Extensive experiments on fine-grained benchmarks (CUB-200-2011 and NABirds) as well as a coarse-grained dataset (CIFAR-10) demonstrate that the proposed framework consistently outperforms or matches state-of-the-art hashing methods, validating its robustness and generalization across different levels of semantic granularity.