Hero-Mamba: Mamba-based Dual Domain Learning for Underwater Image Enhancement
| Authors | Tejeswar Pokuri & Shivarth Rai |
| Year | 2026 |
| Field | Computer Vision |
| arXiv | 2604.16266 |
| Download | |
| Categories | cs.CV |
Abstract
Underwater images often suffer from severe degradation, such as color distortion, low contrast, and blurred details, due to light absorption and scattering in water. While learning-based methods like CNNs and Transformers have shown promise, they face critical limitations: CNNs struggle to model the long-range dependencies needed for non-uniform degradation, and Transformers incur quadratic computational complexity, making them inefficient for high-resolution images. To address these challenges, we propose Hero-Mamba, a novel Mamba-based network that achieves efficient dual-domain learning for underwater image enhancement. Our approach uniquely processes information from both the spatial domain (RGB image) and the spectral domain (FFT components) in parallel. This dual-domain input allows the network to decouple degradation factors, separating color/brightness information from texture/noise. The core of our network utilizes Mamba-based SS2D blocks to capture global receptive fields and long-range dependencies with linear complexity, overcoming the limitations of both CNNs and Transformers. Furthermore, we introduce a ColorFusion block, guided by a background light prior, to restore color information with high fidelity. Extensive experiments on the LSUI and UIEB benchmark datasets demonstrate that Hero-Mamba outperforms state-of-the-art methods. Notably, our model achieves a PSNR of 25.802 and an SSIM of 0.913 on LSUI, validating its superior performance and generalization capabilities.
Engineering Breakdown
Plain English
Hero-Mamba tackles the problem of underwater image enhancement, where images suffer from color distortion, low contrast, and blur due to light absorption and scattering in water. While CNNs and Transformers have been applied to this problem, CNNs can't model long-range dependencies needed for non-uniform degradation patterns, and Transformers become computationally expensive (quadratic complexity) on high-resolution images. The paper proposes Hero-Mamba, a Mamba-based network that processes underwater images in two domains simultaneously: spatial domain (RGB pixels) and spectral domain (FFT frequency components). This dual-domain approach leverages Mamba's linear-time sequential modeling to achieve efficient, high-quality image enhancement without the computational bottlenecks of Transformers.
Core Technical Contribution
The core novelty is the dual-domain learning architecture that processes spatial and spectral information in parallel through a Mamba backbone, rather than treating enhancement as a single-domain problem. Mamba is a state-space model architecture that achieves linear complexity in sequence length, making it fundamentally more efficient than Transformers' quadratic attention while maintaining ability to capture long-range dependencies that CNNs miss. The key insight is that underwater degradation manifests differently in the frequency domain (color shifts, noise distribution) versus spatial domain (detail loss, contrast reduction), so processing both in parallel with selective information fusion allows the network to learn complementary enhancement strategies. This represents a paradigm shift from single-domain CNN/Transformer approaches to a multi-domain architecture optimized for the specific physics of underwater imaging.
How It Works
The architecture takes an underwater RGB image and its FFT-transformed frequency representation as dual inputs, feeding them through separate Mamba-based processing streams. Each stream uses Mamba blocks—which employ selective state-space mechanisms to maintain linear computational complexity while capturing long-range context—to learn domain-specific enhancement transformations. The spatial stream processes raw pixel information to recover details and local contrast, while the spectral stream works on frequency coefficients to correct color distortion and reduce noise artifacts that underwater scattering introduces. The outputs from both streams are fused through learned weighted combination or gating mechanisms, producing an enhanced RGB image that benefits from both spatial detail recovery and spectral color/noise correction. The linear complexity of Mamba (O(N) versus Transformer's O(N²)) allows the model to process high-resolution underwater images efficiently, which is critical since underwater surveys often require high fidelity.
Production Impact
For engineers deploying underwater image enhancement in production systems—autonomous underwater vehicles (AUVs), subsea inspection, marine research pipelines—this approach reduces inference latency and memory footprint compared to Transformer-based methods while maintaining or improving quality. A production system using Hero-Mamba would see significant speed improvements on edge devices or real-time processing scenarios; where a ViT-based approach might require GPU acceleration to process 1080p imagery in acceptable time, Mamba's linear complexity means lower latency and potentially inference on mobile/embedded hardware. The dual-domain design means engineers would need to compute and manage both spatial and spectral representations during inference, adding preprocessing overhead (FFT computation) but typically negligible compared to the model inference itself. Integration would require retraining or fine-tuning on domain-specific underwater datasets (e.g., UIEB, EUVP), and the method assumes access to ground-truth enhanced images or paired degraded-clean datasets for supervised training, which may not be available for all underwater domains.
Limitations and When Not to Use This
The paper's approach assumes underwater degradation can be meaningfully decomposed into spatial and spectral components with learnable separation, which may not hold universally—some degradation types (e.g., severe turbidity in coastal waters) may have tightly coupled spatial-spectral effects where parallel processing offers diminishing returns. While Mamba offers linear complexity, the actual wall-clock speedup depends on implementation maturity and hardware support; as of 2026, Mamba kernels are not as optimized on commodity GPUs as Transformer implementations, potentially negating some theoretical efficiency gains in practice. The method requires paired training data (degraded-enhanced image pairs), which is expensive to obtain for underwater imaging, and performance likely degrades significantly on out-of-distribution underwater environments not represented in training data. The paper appears incomplete (abstract cuts off mid-sentence), so key details about training methodology, benchmark results, and comparative performance against state-of-the-art methods are unavailable for full assessment.
Research Context
This work builds on the emerging use of Mamba and state-space models as alternatives to Transformers in computer vision, following papers like Vision Mamba that showed SSMs could match or exceed ViT performance with better efficiency. It addresses a specific gap in underwater image enhancement, where prior CNN-based methods (like U-Net variants) and recent Transformer approaches (like SwinIR adapted for underwater) have trade-offs between expressiveness and efficiency. The dual-domain learning strategy extends prior work in image restoration (e.g., multi-scale processing, frequency-domain methods) by explicitly using Mamba's architecture to handle both domains efficiently—a design choice uniquely enabled by Mamba's linear scaling. This opens a research direction of applying SSMs to domain-specific restoration tasks where single-domain approaches hit computational or representational limits.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::