Detection-Guided Deep Unfolding for Joint Underwater Image Enhancement and Object Detection

We address the limitation of conventional underwater image enhancement algorithms, which typically prioritize perceptual quality over downstream machine vision requirements. To this end, we first construct the Underwater Joint Enhancement and Detection (UJED) dataset, the first unified benchmark that provides both perceptual reference images and detection annotations within the same domain. Next, we propose Detection-Guided deep unfolding UIE Network (DGU-Net), which integrates physics-guided and detection-guided regularization into a joint optimization framework, balancing visual quality for human perception and detection accuracy for machine vision. We solve the optimization problem iteratively and then unroll the solver into a multistage network, where the optimization variables and regularizers are updated using closed-form solutions and learnable proximal operators. Experimental results demonstrate that the proposed algorithm achieves state-of-the-art detection performance without sacrificing visual quality.