JoVE - Journal of Visualized Experiments
JoVE Visualize
Contact Us

Task-Driven Underwater Image Enhancement via Hierarchical Semantic Refinement

IEEE Transactions on Image Processing : a Publication of the IEEE Signal Processing Society +

|

January 1, 2026

|

January 1, 2026

View abstract on PubMed

Summary

This summary is machine-generated.

This study introduces a new underwater image enhancement (UIE) method that focuses on recovering machine-readable semantics, not just visual appeal. The task-driven framework improves underwater image analysis for marine exploration.

Area Of Science

  • Computer Vision
  • Marine Robotics
  • Image Processing

Background

  • Underwater image enhancement (UIE) is vital for marine exploration but current methods often fail to address semantic corruption.
  • Degradations in underwater images are layer-specific, affecting both shallow and deep features, and entangle with semantic content.
  • Existing UIE methods prioritize perceptual quality, leading to semantic damage that hinders downstream machine vision tasks.

Purpose Of The Study

  • To develop a task-driven UIE framework that redefines enhancement as machine-interpretable semantic recovery.
  • To address the limitations of existing methods in handling irreversible semantic corruption in underwater images.
  • To improve the performance of downstream tasks like segmentation, detection, and saliency analysis on degraded underwater imagery.

Main Methods

  • Proposed a multi-scale underwater distortion-aware generator to identify and prioritize distortions across feature levels.
  • Developed a self-supervised disentanglement strategy using CLIP-based semantic constraints and identity consistency to separate distortions from content.
  • Introduced a task-aware hierarchical enhancement module for refining shallow details and strengthening deep semantics, aligning with machine vision needs.

Main Results

  • The proposed framework effectively recovers machine-friendly semantics from degraded underwater images.
  • Experimental results on segmentation, detection, and saliency tasks demonstrate significant improvements over existing methods.
  • The method successfully compensates for irreversible semantic loss and enhances corrupted content for better machine interpretation.

Conclusions

  • The task-driven UIE framework offers a novel approach to semantic recovery in underwater imagery.
  • This method enhances the robustness and accuracy of marine exploration through improved image analysis.
  • The developed techniques provide a significant advancement for machine vision applications in underwater environments.