【摘要】 针对组合零样本学习中属性与对象概念分离与组合的挑战，提出一种基于多层次特征增强的组合零样本分类网络。该方法利用注意力机制将属性及对象表示为可学习词元，通过跨注意力机制实现概念表示与图像特征的分离和融合。借助预训练的对比语言-图像模型编码提示文本"a photo of[ATTRIBUTE][OBJECT]"，将生成的属性-对象词向量组合嵌入在空间与通道注意力机制下进行特征融合。通过相似度计算完成分类。实验结果表明，所提方法显著提升了模型的泛化能力和性能，在Clothing 16K、UT-Zappos 50K和C-GQA数据集的封闭世界与开放世界测试中均得到有效验证。更多还原

【Abstract】 To address the challenge of separating and combining attribute and object concepts in compositional zero-shot learning, a multi-level feature-enhanced compositional zero-shot classification network was proposed. An attention mechanism was utilized to represent attributes and objects as learnable tokens, and the separation and fusion of conceptual representations and image features were achieved through a cross-attention mechanism. A pre-trained contrastive language-image model was employed to encode the prompt "a photo of [ATTRIBUTE][OBJECT]", and the generated attribute-object word vector compositions were fused under spatial and channel attention mechanisms. Classification was accomplished through similarity calculation. Experimental results demonstrate that the proposed method significantly enhances the model’s generalization capability and performance, validated effectively in both closed-world and open-world tests on the Clothing 16K, UT-Zappos 50K, and CGQA datasets.更多还原