跳转至

面向细粒度广义类别发现的傅里叶自监督方法

文章背景与核心概要

广义类别发现（GCD）旨在从未标注的数据中识别已知类别并发现未知类别。传统的自监督和对比学习方法往往依赖表层的视觉特征，在处理细粒度区分类别时显得力不从心。本文介绍了一种名为“傅里叶自监督”（Fourier Self-Supervision）的新型双频滤波方法，旨在解决这一痛点。

该技术利用图像的傅里叶变换，通过低通滤波器捕获广泛的高层类别特征，同时通过高通滤波器捕捉边缘和纹理等精细细节。这些滤波器在专用且重叠的潜在空间中运行，显著增强了模型的判别能力。在目标类别数量未知的情况下，该方法在多个细粒度数据集上的表现依然超越了现有的最先进模型。

摘要 (Summary)

傅里叶自监督（Fourier Self-Supervision）为广义类别发现（GCD）引入了一种新颖的双频滤波方法。传统的自监督方法往往依赖于表层的视觉线索，难以进行细粒度的区分。该方法利用低通滤波器来捕获广泛的高层类别特征，并利用高通滤波器来捕获精细细节（如边缘和纹理）。通过在专用和重叠的潜在空间中运行，这种双重方法显著增强了判别能力——即使在目标类别数量未知的情况下，也能在细粒度数据集上表现优于最先进的模型。

Fourier Self-Supervision introduces a novel dual-frequency filtering approach for Generalized Category Discovery (GCD). Traditional self-supervised methods often rely on superficial visual cues, struggling with fine-grained distinctions. This method uses a low-pass filter to capture broad, high-level category traits and a high-pass filter to capture fine details (such as edges and textures). Operating across dedicated and overlapping latent spaces, this dual approach significantly enhances discriminative power—outperforming state-of-the-art models on fine-grained datasets even when the number of target classes is unknown.

论文元数据 (Paper Metadata)

  • arXiv 标识符： arXiv:2608.08963 [cs.CV]
  • 作者： Sarah Rastegar, Mina Ghadimi Atigh, Pascal Mettes, Yuki M. Asano, Cees G. M. Snoek
  • 提交日期： 2026年8月9日
  • 主要学科： 计算机视觉与模式识别 (cs.CV)
  • 收录会议： ECCV 2026
  • 代码仓库： GitHub - FourEx
  • arXiv Identifier: arXiv:2608.08963 [cs.CV]
  • Authors: Sarah Rastegar, Mina Ghadimi Atigh, Pascal Mettes, Yuki M. Asano, Cees G. M. Snoek
  • Submission Date: August 9, 2026
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Accepted Venue: ECCV 2026
  • Code Repository: GitHub - FourEx

摘要全文 (Abstract)

广义类别发现旨在识别已知类别，同时在未标注数据中识别出新类别。现有方法通常基于自监督和对比学习，往往难以捕捉细粒度的区别，更多地依赖于表层视觉线索，而不是人类用于分类的内在属性。

我们引入了傅里叶自监督（Fourier Self-Supervision），它利用图像的傅里叶变换来增强对细微差异的区分，并支持新类别的发现。我们的方法采用双频滤波策略： 1. 低通滤波器： 首先提取捕捉高层类别信息的广泛、抽象属性。 2. 高通滤波器： 强调对细粒度识别至关重要的边缘和纹理等精细细节。

每个滤波器都在专用的潜在空间上运行，它们重叠的表示共同产生了一个更丰富、更完整的特征空间。这种双频方法不仅改进了特征提取以识别新类别，还增强了模型在细粒度类别发现中的判别能力。在多个细粒度数据集上的实验表明，结合傅里叶自监督的方法优于最先进的方法，即使在类别数量未知的情况下也是如此，这证明了其在广义类别发现中的有效性。

Generalized Category Discovery aims to recognize known categories while identifying novel ones within unlabeled data. Existing methods, typically based on self-supervision and contrastive learning, often struggle to capture fine-grained distinctions, relying on superficial visual cues rather than the intrinsic attributes humans use for categorization.

We introduce Fourier Self-Supervision, which leverages the Fourier transform of images to enhance the discrimination of subtle differences and support the discovery of new categories. Our method employs a dual frequency filtering strategy: 1. Low-pass filter: First extracts broad, abstract attributes that capture high-level category information. 2. High-pass filter: Emphasizes fine details such as edges and textures that are essential for fine-grained recognition.

Each filter operates on a dedicated latent space, and their overlapping representations together yield a richer, more complete feature space. This dual-frequency approach not only refines feature extraction to identify novel categories, but also strengthens the model's discriminative power in fine-grained category discovery. Experiments on multiple fine-grained datasets show that incorporating Fourier Self-Supervision outperforms state-of-the-art methods, even when the number of classes is unknown, demonstrating its effectiveness for Generalized Category Discovery.