PLUM: A Prompt-Guided Lightweight Unified Model for Enhancement of Multidegraded Underwater Image

Abstract

Degradation in underwater images is commonly caused by haze, low contrast, blur, and color distortion, which can hinder the effectiveness of many computer vision tasks. Most of the existing image enhancement works treat these degradations separately, thus limiting their applicability. To handle this problem, we propose a unified model to learn multiple degradations concurrently, improving resilience to practical conditions. First, we generate various degraded versions (haze, blur, low-contrast, and color distortion) of the input degraded image through domain translation. Next, original degraded and translated images are fed into a novel multidegraded enhancement network, which employs multiple encoders with an intermediate feature-swapping mechanism and a single decoder. Each encoder block enhances the network’s ability to characterize internal relationships in spatial and frequency domains by incorporating the fast Fourier transform. Intermediate feature swapping encourages the model to learn more robust representations. The decoder block contains the proposed prompt-guided spatial-frequency cross-attention block to encode degraded visual information effectively. Extensive experiments on several underwater image enhancement data sets demonstrate that our model outperforms the state-of-the-art methods with less computational cost (e.g., 1.20% structural similarity index boost on UIEB data set with 84.1% less parameter).