Scholarlink

<p>Frequency-based Neural networks have attracted considerable interest due to their global receptive fields and advantageous time efficiency. Recent research has largely utilized global frequency filters within frequency-based networks for the extraction of image features. These global frequency filters are complex tensors identical in size to the feature itself, necessitating a predefined size during the training phase. Altering the size post-training mandates retraining, as merely interpolating filters fails to preserve the continuous nature of both the real-imaginary and magnitude-phase relationships. Such interpolation undermines the filter’s Practical in feature extraction, posing a challenge for multi-size image feature extraction in conventional frequency-domain networks. To address this challenge, we introduce the Meta Global Filter Network (MGFNet), a novel framework that replaces static frequency filters with dynamic filter generators. Given a inference image size, the generators synthesizes corresponding frequency-domain filters while ensuring consistent frequency-domain characteristics across scales, thereby enabling effective feature extraction from images of varying sizes. Furthermore, we investigate and propose three distinct methodologies for the filter generator mechanism: 1) an Implicit Filter Generator (IFG) implemented via neural networks, 2) an Explicit Filter Generator (EFG) utilizing the band-based filter bias, and 3) a Convolution Theorem Generator (CTG) leveraging the convolution theorem. These generators are designed from complementary perspectives and are employed synergistically to enhance the expressive power of the resulting frequency filters. Finally, we implement a frequency-domain re-parameterization method. During inference, MGFNet pre-computes and consolidates the outputs from the three filter generators into a single, unified global frequency-domain filter, significantly boosting computational efficiency and reducing inference latency. Extensive experiments conducted on the ImageNet, MSCOCO, and nuScenes datasets demonstrate that our model achieves state-of-the-art performance across various datasets when compared to frequency networks. The code is available on <a href="https://github.com/WallelWan/MGFNet">https://github.com/WallelWan/MGFNet</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

MGFNet: Meta Global Filter Network for multi-size image feature extraction

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Hong Zhang,
  • Jiaxu Wan,
  • Jianbo Song,
  • Hanyang Liu,
  • Ding Yuan,
  • Yifan Yang

摘要

Frequency-based Neural networks have attracted considerable interest due to their global receptive fields and advantageous time efficiency. Recent research has largely utilized global frequency filters within frequency-based networks for the extraction of image features. These global frequency filters are complex tensors identical in size to the feature itself, necessitating a predefined size during the training phase. Altering the size post-training mandates retraining, as merely interpolating filters fails to preserve the continuous nature of both the real-imaginary and magnitude-phase relationships. Such interpolation undermines the filter’s Practical in feature extraction, posing a challenge for multi-size image feature extraction in conventional frequency-domain networks. To address this challenge, we introduce the Meta Global Filter Network (MGFNet), a novel framework that replaces static frequency filters with dynamic filter generators. Given a inference image size, the generators synthesizes corresponding frequency-domain filters while ensuring consistent frequency-domain characteristics across scales, thereby enabling effective feature extraction from images of varying sizes. Furthermore, we investigate and propose three distinct methodologies for the filter generator mechanism: 1) an Implicit Filter Generator (IFG) implemented via neural networks, 2) an Explicit Filter Generator (EFG) utilizing the band-based filter bias, and 3) a Convolution Theorem Generator (CTG) leveraging the convolution theorem. These generators are designed from complementary perspectives and are employed synergistically to enhance the expressive power of the resulting frequency filters. Finally, we implement a frequency-domain re-parameterization method. During inference, MGFNet pre-computes and consolidates the outputs from the three filter generators into a single, unified global frequency-domain filter, significantly boosting computational efficiency and reducing inference latency. Extensive experiments conducted on the ImageNet, MSCOCO, and nuScenes datasets demonstrate that our model achieves state-of-the-art performance across various datasets when compared to frequency networks. The code is available on https://github.com/WallelWan/MGFNet.