RGI-Net: A Recursive Gated Injection Network for Underwater Image Restoration in Marine Inspection

Abstract

Background: Underwater optical imagery used in marine inspection is degraded by wavelength-dependent attenuation and scattering, which couple color cast with structural haze. Existing restorers either treat both degradations in one stage or cascade separate networks and exchange information only through reconstructed images. Methods: This paper proposes RGI-Net, combining (i) a shared-weight two-pass pipeline that disentangles Lab chromatic correction from RGB structure refinement without duplicating the backbone and (ii) a 1×1 Gated Injection Bridge that transfers Pass-1 bottleneck features into Pass 2 under a learned element-wise gate, together with a recursive contrastive objective that includes a cross-pass consistency term. All methods are compared on identical test partitions under one evaluation pipeline (UIEB train/test, LSUI, EUVP, U45, C60). Results: Against WaterNet, PUIE-Net, CCL-Net, and Phaseformer, RGI-Net attains the best UIEB UIQM (2.21) and UCIQE (0.372), with PSNR/SSIM statistically indistinguishable from Phaseformer (Wilcoxon p>0.5). On the reported GPU, wall-clock latency is 41.00 ms, versus 61.64 ms for Phaseformer at 256×256, while RGI-Net has the largest parameter count and FLOPs in the comparison. Conclusions: Shared two-pass decoupling with gated bottleneck injection improves UIEB no-reference indices. Lower latency than Phaseformer on one workstation GPU is not a claim of overall computational efficiency, given the larger parameter and FLOP counts. UIQM and UCIQE are operational scores, not perceptual or inspection-task evidence; on EUVP, the best UIQM/UCIQE coincide with the lowest PSNR/SSIM among compared methods, which remains a substantial limitation.