کنفرانس بین المللی فناوری اطلاعات و دانش

English

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Prompt-Based Composed Fashion Image Retrieval via Gated Detail-Enhanced Dual Cross-Attention Difference Modeling

نویسندگان :

Kosar Keshavarz¹ Reza Azmi²

1- دانشگاه الزهرا(س) 2- دانشگاه الزهرا(س)

کلمات کلیدی :

Composed image retrieval،Composed query،Contrastive learning،Fashion retrieval،Multimodal retrieval،Text-guided image retrieval

چکیده :

With the rapid growth of online shopping and the vast amount of fashion-related visual content on the internet, accurate methods for fashion image retrieval have become increasingly important to enhance user satisfaction. The fashion domain is inherently fine-grained, characterized by subtle details such as color, pattern, cut, and embellishments, where even small variations lead to distinct styles. To address the limitations of purely text-based or image-based queries, we adopt a text-guided retrieval approach in which a reference image and a natural-language description jointly define the user’s intent. This paper extends sentence-level prompt-based retrieval frameworks by introducing explicit image-difference modeling. The proposed Gated Detail-Enhanced Dual Cross-Attention (GDD-CA) module models the relationship between reference and target images through dual cross-attention and a gated detail-enhancement mechanism, enabling the network to capture subtle, fine-grained visual variations. Experimental results on the Fashion-IQ dataset demonstrate that integrating detail-enhanced image-difference modeling into the prompt-based structure improves retrieval performance, achieving a 1.14% gain in Recall over previous methods.