Scholarlink

<p>This paper proposes a novel neural network architecture that integrates wavelet transform principles to achieve efficient and effective feature learning. The core innovation lies in two specialized modules: 1) The Dual Frequency Branch Module (DFBM) performs wavelet-like frequency decomposition to separate input features into distinct high- and low-frequency components. Specifically, it employs a variance-based modulation mechanism to enhance low-frequency components, while utilizing convolutional operations combined with dimension-wise transformations to refine high-frequency details. 2) The Cross-Axis Feature Refiner (CAFR) implements a computationally efficient enhancement strategy through a clever sequence of lightweight convolutional layers, enabling cross-axis feature interaction with minimal computational overhead. Extensive evaluations across multiple benchmark datasets demonstrate that our approach maintains highly competitive performance, exhibiting a performance gap of less than 10 percent compared to state-of-the-art methods. Crucially, the architecture achieves this with substantially reduced resource demands: it requires fewer than 50 percent of the FLOPs and consumes only 30 percent of the GPU memory compared to the best-performing baseline models. The code of this work is publicly available at <a href="https://github.com/Lynxsx/DBFM">https://github.com/Lynxsx/DBFM</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

Learned image compression with dual frequency branch modulation

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Lingchen Qiu,
  • Enjian Bai,
  • Yun Wu,
  • Yuwen Cao

摘要

This paper proposes a novel neural network architecture that integrates wavelet transform principles to achieve efficient and effective feature learning. The core innovation lies in two specialized modules: 1) The Dual Frequency Branch Module (DFBM) performs wavelet-like frequency decomposition to separate input features into distinct high- and low-frequency components. Specifically, it employs a variance-based modulation mechanism to enhance low-frequency components, while utilizing convolutional operations combined with dimension-wise transformations to refine high-frequency details. 2) The Cross-Axis Feature Refiner (CAFR) implements a computationally efficient enhancement strategy through a clever sequence of lightweight convolutional layers, enabling cross-axis feature interaction with minimal computational overhead. Extensive evaluations across multiple benchmark datasets demonstrate that our approach maintains highly competitive performance, exhibiting a performance gap of less than 10 percent compared to state-of-the-art methods. Crucially, the architecture achieves this with substantially reduced resource demands: it requires fewer than 50 percent of the FLOPs and consumes only 30 percent of the GPU memory compared to the best-performing baseline models. The code of this work is publicly available at https://github.com/Lynxsx/DBFM.