Li, Hongjun, Wang, Nan, Chen, Anqing, Liu, Jiang, Ma, Wanli, Liu, Weide, Ju, Yakun, Rosin, Paul L.  ORCID: https://orcid.org/0000-0002-4965-3884, Liu, Hantao ORCID: https://orcid.org/0000-0003-4544-3481 and Zhou, Wei
2026.
EHIN: Early-aware hierarchical interaction network for weakly-supervised referring image segmentation.
Neurocomputing
659
, 131764.
10.1016/j.neucom.2025.131764
|
Abstract
Referring image segmentation (RIS) aims to segment target regions in images based on natural language descriptions. Although weakly-supervised RIS frameworks have been proposed to reduce reliance on costly manual annotations, their performance remains limited due to both the low quality of pseudo-labels and the inherent difficulty in achieving effective interaction between visual and textual features. In this paper, we propose a novel weakly-supervised framework named Early-aware Hierarchical Interaction Network (EHIN). The proposed network includes two key components, which are designed to enhance pseudo-label generation and improve the interaction between visual and textual features for RIS, respectively. First, EHIN incorporates an Early-aware Contrastive Learning Module (ECLM) that enhances feature discrimination by leveraging contrastive learning to distinguish target features from background noise. By integrating the module early into the processing pipeline, ECLM operates on raw image features directly, preserving richer visual details while reducing reliance on labeled data and thus improving the reliability of pseudo-labels. Second, EHIN integrates a Hierarchical Interaction Prompt Module (HIPM) to facilitate comprehensive interaction between visual and textual features and enhance subsequent feature fusion. Extensive experimental results on four benchmark datasets demonstrate that the proposed EHIN outperforms the state-of-the-art RIS. Code is available at https://github.com/CDUT-DBGroup/MFP-TRIS.
| Item Type: | Article |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Computer Science & Informatics |
| Publisher: | Elsevier |
| ISSN: | 0925-2312 |
| Date of Acceptance: | 8 October 2025 |
| Last Modified: | 04 Nov 2025 11:01 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/182058 |
Actions (repository staff only)
 |
Edit Item |
Altmetric
Altmetric