Perception-driven underwater image enhancement via multi-level feature fusion

In recent years, underwater image enhancement technology has attracted increasing attention due to its significant contribution to high-level visual tasks. However, existing methods fail to balance enhancement performance and computational cost, limiting their practical application. To address this issue, we propose a perception-driven underwater image enhancement framework based on multi-level feature fusion. Specifically, we employ depthwise and pointwise convolutions to build a lightweight backbone. To compensate for the limitations of the lightweight network in feature extraction, we design a multi-scale texture-structure auxiliary network to extract texture and structure features at different levels, and incorporate these features as perceptual information into the backbone network. Furthermore, to effectively mitigate the color distortion and low contrast in degraded images, we propose a histogram distribution loss function and an adaptive hybrid color space loss function. Extensive experiments demonstrate that the proposed perception-driven framework outperforms existing state-of-the-art methods, achieving superior enhancement quality at lower computational costs. Additionally, underwater object detection experiments validate that our method significantly improves performance in high-level visual tasks. The code is available at https://github.com//PDMF.