GHGL-Net: gated hierarchical group learning with decoupled spectral-spatial dual attention for underwater image restoration

Abstract The underwater optical imaging is degraded by two different phenomena, namely the molecular absorption of light, which is selective in the wavelength of the light, and the anisotropic scattering of light by particles in the Mie-regime. Combined together these result in colour distortion according to depth, loss in contrast and loss of texture. These processes are in operation at very different spatial scales, and thus require different correction methods. Unfortunately, the current restoration approaches (based on convolutional and Transformer models) combine these two processes in the same feature representations resulting in oversmoothed restorations, loss of detail in edges and colours, which are not very reliable. This paper introduces a deep restoration neural network designed following four physics-driven principles, called GHGL-Net. (i) The Multi-Branch Dynamic Fusion Stem extracts the features simultaneously by 3 branches ( $$3\times \! 3$$ , $$5\! \times \! 5$$ , and $$7\! \times \! 7$$ ). They are fused through a learned channel-wise attention weighting function and fine-grained scattering loss and global colour bias are all encoded at the same time across complementary receptive fields. (ii) A Hierarchical Gating Mechanism computes a channel-wise global context vector at each network stage via adaptive average pooling. It multiplicatively modulates Group Structure outputs prior to feature re-injection, mitigating the contrast washout caused by unconditional additive skip connections. (iii) Decoupled Dual-Attention Blocks perform independent channel-level spectral recalibration and pixel-level spatial masking in parallel, enabling specialised convergence toward each physical restoration objective. (iv) A Residual Dense Block at the reconstruction head provides dense inter-layer feature reuse, supervised by a joint $$\mathcal {L}_{1}$$ -VGG-19 perceptual objective that enforces both pixel accuracy and perceptual naturalness. Evaluated on three public benchmarks, GHGL-Net attains the best full-reference performance among all compared methods. Ablation studies confirm the statistically significant and non-redundant contribution of each component. Improved restoration quality of this kind is potentially beneficial for downstream applications such as autonomous underwater vehicle (AUV) navigation, marine biology surveys, and offshore pipeline inspection systems, though this manuscript does not evaluate these tasks directly.