A fashion image retrieval with hybrid features

Liren Tian; Xiaoyun Zhang; Taiqi Liu; Pinjie He; Zhigao Zeng

doi:10.1117/12.3045417

15 January 2025 A fashion image retrieval with hybrid features

Liren Tian, Xiaoyun Zhang, Taiqi Liu, Pinjie He, Zhigao Zeng

Author Affiliations +

Proceedings Volume 13513, The International Conference Optoelectronic Information and Optical Engineering (OIOE2024); 135130V (2025) https://doi.org/10.1117/12.3045417
Event: The International Conference Optoelectronic Information and Optical Engineering (OIOE2024), 2024, Wuhan, China

Abstract

Fashion image conditional retrieval is a method that utilizes textual feedback to guide the modification of fashion image content and uses it as retrieval query condition. The key point of this task is effectively integrating the semantic spaces of fashion images and text to support subsequent retrieval tasks. This paper proposes a hybrid feature cross-attention fashion image retrieval approach. The approach first extracts comprehensive fashion image features and text features using a pre-trained CLIP model and then uses a multi-head attention mechanism for semantic enhancement. Next, a hybrid feature cross-attention mechanism is utilized to obtain hybrid feature representations of images and text, which are fused into a single feature vector to retain more semantic information from both modalities. Finally, a two-stage training approach using contrastive learning is applied to train the network. Experimental results demonstrate the approach's good performance on various fashion image conditional retrieval metrics on the FashionIQ dataset and Shoes dataset.

(2025) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Liren Tian, Xiaoyun Zhang, Taiqi Liu, Pinjie He, and Zhigao Zeng "A fashion image retrieval with hybrid features", Proc. SPIE 13513, The International Conference Optoelectronic Information and Optical Engineering (OIOE2024), 135130V (15 January 2025); https://doi.org/10.1117/12.3045417

Access the abstract

PROCEEDINGS
8 PAGES

DOWNLOAD PAPER SAVE TO MY LIBRARY