跨越表达鸿沟：面向电子商务的零样本组合图像检索(ZS-CIR)技术研究与应用

Bridging the Expression Gap: Research and Applications of Zero-Shot Composed Image Retrieval for E-Commerce

摘要: 为解决电子商务搜索中用户“视觉特征保留 + 语义属性修改”的复合查询需求与传统文本搜索表达局限形成的“表达鸿沟”，本文聚焦零样本组合图像检索(ZS-CIR)技术，通过构建统一数学框架，系统性梳理其技术路径、阐释自掩码投影与噪声注入等关键方法的理论依据，并在典型数据集上对文本反演、纯语言训练与合成数据驱动等主流范式开展全面评测与对比。实验结果显示，纯语言训练方法在极低训练成本下实现实用性能(Recall@10 38.5%，推理延迟18 ms)，验证了语言空间模拟视觉修改的可行性；合成数据方法依托规模效应达成当前最优性能(Recall@10 46.8%)。本文从技术图谱、理论支撑与系统架构层面，为ZS-CIR在电商场景的研究与应用提供系统参考。

Abstract: To address the “expression gap” between users’ composite query needs of “visual feature retention + semantic attribute modification” and the expressive limitations of traditional text search in e-commerce, this paper focuses on Zero-Shot Composed Image Retrieval (ZS-CIR) technology. By constructing a unified mathematical framework, it systematically sorts out technical pathways, clarifies the theoretical basis of key methods such as Self-Masking Projection (SMP) and noise injection, and conducts comprehensive evaluations and comparisons of mainstream paradigms including textual inversion, language-only training, and synthetic data-driven approaches on benchmark datasets. Experimental results show that the language-only training method achieves practical performance with minimal training cost (Recall@10 38.5%, inference latency 18 ms), verifying the feasibility of simulating visual modifications in the linguistic space; the synthetic data-driven method attains state-of-the-art performance (Recall@10 46.8%) through scaling effects. From the perspectives of technology mapping, theoretical underpinnings, and system architecture, this paper provides a systematic reference for the research and application of ZS-CIR in e-commerce scenarios.