| 摘 要: 图像分类任务常涉及层次化类别结构、非均匀语义差异及细粒度区分,这些特征在欧几里得空间中难以有效建模。现有方法多假设各向同性曲率,并依赖基于距离的softmax分类器,难以捕捉特征方向变化和形成紧凑类簇。为解决上述问题,提出各向异性双曲神经网络(AHNN),以Vision Transformer为特征骨干,在指数映射前引入可学习各向异性变换,实现方向特定扩展,更好对齐视觉数据内在结构。同时设计双曲质心损失(Hyperbolic Centroid Loss),结合基于双曲距离的交叉熵、原型正则化和辅助欧几里得损失,提升类内聚类与训练稳定性。在六个有代表性基准数据集上的实验表明,所提方法显著优于欧几里得基线和现有双曲模型,尤其在细粒度和层次化数据集上优势明显。 |
| 关键词: 双曲几何 各向异性嵌入 Vision Transformer 度量学习 基于原型的分类 层次表示 |
|
中图分类号: TP391.4
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Anisotropic Hyperbolic Neural Networks for Fine-grained and Hierarchical Image Classification |
|
CHEN Yi1,2,3,4, MA Lingkun1,2,3,4, ZHANG Weichuan1,2,3,4
|
1.College of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi'2.'3.an 710021;4.China
|
| Abstract: Image classification tasks often involve hierarchical category structures, non-uniform semantic differences, and fine-grained distinctions, which are difficult to effectively model in Euclidean space. Existing methods often assume isotropic curvature and rely on distance-based softmax classifiers, making it difficult to capture changes in feature direction and form compact class clusters. To address these issues, we propose the Anisotropic Hyperbolic Neural Network (AHNN), which utilizes the Vision Transformer as the feature backbone and introduces a learnable anisotropic transformation before exponential mapping to achieve direction-specific expansion and better align with the inherent structure of visual data. Additionally, we design the Hyperbolic Centroid Loss, which combines cross-entropy based on hyperbolic distance, prototype regularization, and auxiliary Euclidean loss to enhance intra-class clustering and training stability. Experiments on six representative benchmark datasets demonstrate that the proposed method significantly outperforms Euclidean baselines and existing hyperbolic models, especially on fine-grained and hierarchical datasets. |
| Keywords: hyperbolic geometry anisotropic embedding Vision Transformer metric learning prototype-based classification hierarchical representation |
