Scholarlink

<p>With the widespread application of remote sensing technology, the number of remote sensing images has shown explosive growth. However, traditional image retrieval methods often encounter computational challenges posed by high-dimensional feature spaces. In order to address this issue, the deep hashing method has gradually become a mainstream method due to its efficient retrieval and low storage cost. However, deep hashing methods based on Convolutional Neural Networks (CNNs) focus on local features, failing to effectively capture global semantic information, and leading to a blurring of the correlations between image features. Furthermore, most existing hashing methods construct training batches by random sampling or semi-hard negative mining, which causes sample imbalance, thus biasing the training process towards majority categories and degrading the performance of the model. To remedy the aforementioned issues, we propose a novel deep remote sensing hashing framework, called Improved Hierarchical Transformer-Based Multi-Similarity Hashing via Label-guided learning, to learn similarity-preserving and discriminative binary descriptors. Specifically, to simultaneously capture global and local semantic information, the simplified and improved hierarchical transformer framework is designed as the feature learning module, which could achieve the powerful ability to feature extraction of more complex networks. By introducing pair mining and weighting operations to jointly calculate relative similarity and self-similarity between images, the multi-similarity mining strategy is developed to alleviate the impact of sample imbalance in the training process. Furthermore, the class-specific supervisory signals from the label-guided learning are utilized to effectively guide the parameter optimization of deep hashing. Comprehensive experiments on three datasets demonstrate that our deep hashing approach can achieve optimal performance for remote sensing image retrieval. The source code is available at <a href="https://github.com/QinLab-WFU/HTMSH.">https://github.com/QinLab-WFU/HTMSH.</a></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

Improved hierarchical transformer-based multi-similarity hashing via label-guided learning for remote sensing image retrieval

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Xiaorong Chen,
  • Jinkui Hou,
  • Jiangyan Dai,
  • Lei Huang,
  • Wenfeng Zhang,
  • Qibing Qin

摘要

With the widespread application of remote sensing technology, the number of remote sensing images has shown explosive growth. However, traditional image retrieval methods often encounter computational challenges posed by high-dimensional feature spaces. In order to address this issue, the deep hashing method has gradually become a mainstream method due to its efficient retrieval and low storage cost. However, deep hashing methods based on Convolutional Neural Networks (CNNs) focus on local features, failing to effectively capture global semantic information, and leading to a blurring of the correlations between image features. Furthermore, most existing hashing methods construct training batches by random sampling or semi-hard negative mining, which causes sample imbalance, thus biasing the training process towards majority categories and degrading the performance of the model. To remedy the aforementioned issues, we propose a novel deep remote sensing hashing framework, called Improved Hierarchical Transformer-Based Multi-Similarity Hashing via Label-guided learning, to learn similarity-preserving and discriminative binary descriptors. Specifically, to simultaneously capture global and local semantic information, the simplified and improved hierarchical transformer framework is designed as the feature learning module, which could achieve the powerful ability to feature extraction of more complex networks. By introducing pair mining and weighting operations to jointly calculate relative similarity and self-similarity between images, the multi-similarity mining strategy is developed to alleviate the impact of sample imbalance in the training process. Furthermore, the class-specific supervisory signals from the label-guided learning are utilized to effectively guide the parameter optimization of deep hashing. Comprehensive experiments on three datasets demonstrate that our deep hashing approach can achieve optimal performance for remote sensing image retrieval. The source code is available at https://github.com/QinLab-WFU/HTMSH.