NCF-Net: Noise Corruption Fusion Network for Robust and Perception-Consistent Underwater Image Enhancement

Underwater images often suffer from multiple degradations, such as color distortion, detail loss, and contrast reduction, which significantly impact downstream visual tasks. Existing methods struggle with insufficient degradation modeling and weak activation of latent features, especially under coupled distortions. To address these challenges, we propose the Noisy Corruption Fusion Network (NCF-Net), a novel underwater image enhancement framework. NCF-Net introduces Gaussian noise perturbations to disrupt the original feature distribution, thereby activating structurally informative yet low-response regions. A noise-aware feature fusion strategy is employed to construct a more expressive representation. To improve multi-scale degradation adaptability, we design a Noise Feature Extractor (NFE) that integrates multi-branch convolutions with instance normalization and channel attention for fine-grained feature aggregation. Additionally, a coordinate attention mechanism is embedded in the encoder-decoder to enhance the joint modeling of spatial and semantic dependencies. Extensive experiments on the UIEB, EUVP, and U45 datasets show that NCF-Net consistently outperforms state-of-the-art methods in terms of PSNR, SSIM, and UIQM. Furthermore, the proposed method demonstrates superior robustness and generalization in downstream tasks such as feature matching and saliency detection.