SFP-Net: Scatter-feature-perception network for underwater image enhancement

Abstract

Inspired by atmospheric scattering light models, substantial progresses have been achieved in deep learning-based methods for underwater image enhancement. However, these methods suffer from a deficiency in accurately modeling scattering information, which can incur some quality issues of visual perception. Moreover, insufficient attention to the key scene features leads to enhanced images that lack fine-grained information. To alleviate these challenges, we propose an efficient scatter-feature-perception network(SFP-Net). It consists of two core ideas: firstly, the dark channel map is synergistically combined with the K-value map to precisely perceive scattering light features within the scene. Subsequently, multi-scale cross-space learning is used to capture the inter-dependencies between channels and spatial positions, facilitating the perception of scene feature information. Besides, the adaptive scatter feature loss is formulated on the basis of the atmospheric scattering model, which evaluates the impact of scattered light. Extensive experimental results demonstrate that our model effectively mitigates the influence of underwater environmental factors, circumvents interference caused by image depth of field, and exhibits superior performance in terms of adaptability and reliability. Notably, our model achieves maximum values of 29.76 and 0.91 on the PSNR and SSIM metrics, which indicates superior enhancement effects compared to existing methods. Meanwhile, the UCIQE and UIQM metrics also reached 0.431 and 2.763 respectively, which are more in line with human visual preferences.