Underwater Image Enhancement via Structure–Texture Disentanglement

Ge, H., Li, W., Zhang, Q., Zhang, T., Wang, D., Fu, B. and Auat Cheein, F. (2026) Underwater Image Enhancement via Structure–Texture Disentanglement. Signal Processing: Image Communication, 149. ISSN 09235965

Full text not available from this repository.

Abstract

Underwater Image Enhancement (UIE) seeks to recover high-quality images with sharp structures and rich textures from degraded images affected by low contrast, blur, and color cast, etc. Significant progress has been achieved in this field, yielding promising results. However, existing most UIE methods often process multi-scale features indiscriminately and neglect the distinct representations of shallow features for textural information and deep features for structural information in image, resulting in blurred structures and smoothed textures in enhanced underwater images. To address these issues, we introduce a Structure–Texture Disentangled Network for UIE, termed STDN. Specifically, we first leverage a pretrained feature extractor on large-scale high-quality images to establish a latent undisturbed underwater feature space. To decouple and enhance structural and textural features from the feature space, we introduce a Residual Channel-guided Structure Extraction (RCSE) module and an Interactive Attention-guided Detail Preservation (IADP) module. RCSE removes illumination and haze-like effect via residual channel extraction, guiding deep structural feature decoupling. IADP uses a dual-branch squeeze-excitation mechanism to enhance shallow feature interaction and preserve richer textures. Following that, to adaptively reconstruct these features, a Multi-scale Dynamic Feature Aggregation (MDFA) module is developed to assign adaptive weights to different feature scales. In addition, we also devise an Inner–Outer Nested Pyramid Decoder (IONPD), enabling a precise feature reconstruction through feature decoding from both intra-scale and inter-scale perspectives. Extensive experiments demonstrate the superiority of our STDN in UIE, as evidenced by its effectiveness in downstream salient object detection. Our code is publicly available at https://github.com/MXINPRIME/STDN

Item Type:	Article
Additional Information:	Full text not available from this repository
Keywords:	Underwater image enhancement, Structure–texture disentangling, Dynamic feature aggregation
Divisions:	Departments > Engineering Research Centres > Harper Institute of Technology
Depositing User:	Miss Anna Cope
Date Deposited:	16 Sep 2026 08:07
Last Modified:	16 Sep 2026 08:07
URI:	https://hau.repository.guildhe.ac.uk/id/eprint/18424

Actions (login required)

Edit Item