CITG: A Cascade in-Situ and Task-Guided Enhancement Method for Improving Underwater Image Quality

The physical properties of water, such as absorption and scattering, induce severe color distortion, low contrast, and detail blur in underwater images, posing a significant challenge for visual analysis. To address this, we propose a Cascaded In-situ and Task-guided Enhancement (CITG) model. CITG employs a two-stage strategy to progressively enhance degraded underwater images. The first stage, In-situ Enhancement (ISE), leverages inherent scene-specific priors to perform preliminary restoration. Using a multi-scale enhancement architecture, ISE adaptively corrects color imbalances and enhances structural details, recovering fundamental visual characteristics. The output of ISE provides scene-adaptive constraints that guide the second stage, narrowing the solution space and preventing unrealistic artifacts. The second stage, Task-Perceptual Guidance (TPG), integrates twin adversarial contrastive learning with a task-aware feedback module. By leveraging gradient information from a downstream vision task network, TPG dynamically identifies critical regions, such as object edges and textural features. Extensive experiments on the URPC2020 and VDD datasets demonstrate that CITG outperforms compared methods in quantitative metrics (UIQM, UCIQE) while significantly enhancing target edge information, validating its superiority in underwater image enhancement.