FCUnet: An Underwater Image Enhancement Hybrid Network via Fused Feature‐Guided Cross‐Attention

Underwater images captured by optical vision systems often suffer from degradation due to the absorption and scattering of reflected light. To mitigate the impact of the complex underwater environment on imaging quality, data‐driven methods for underwater image enhancement have emerged as an effective strategy. However, convolutional neural networks exhibit limitations in processing global information interactions due to their inductive biases. Concurrently, transformers often overlook local details. To address these challenges, we propose FCUnet, a hybrid network designed to enhance the quality of underwater images. First, we propose a colour deviation preprocessing module that incorporates a multi‐scale channel‐wise attention for the fusion feature map to capture differences in rich spatial hierarchical information across channels. Second, we design the cross‐attention block guided by multiple features to efficiently acquire local fine details and global scene characteristics. Finally, a universal feature fusion unit and a colour loss term are proposed to enhance the network's sensitivity to colour and texture deviation information. Ablation studies confirm the effectiveness of each individual component. Experimental results on public benchmark datasets demonstrate that the proposed FCUnet outperforms other competitive methods, achieving superior performance in both qualitative and quantitative evaluations.