Abstract
Underwater image enhancement is critical for marine exploration but faces significant challenges due to light attenuation and scattering. We present USWformer, a lightweight Transformer-based model that integrates Discrete Wavelet Transforms (DWT) directly into the attention mechanism. By leveraging frequency-domain sparsity, USWformer achieves state-of-the-art performance with only 1.19 million parameters, significantly outperforming much larger models in both computational efficiency and restoration quality.
Introduction
Capturing clear underwater images is hindered by low contrast, heavy color tints, and blur caused by light scattering. While deep learning models, particularly Transformers, have advanced the field, they often suffer from high computational redundancy and a failure to exploit frequency-domain characteristics that distinguish image structure from noise. USWformer addresses these limitations by introducing a sparse wavelet-based architecture that is both computationally lean and effective at "seeing" through murky water.
Methodology
The core innovation of USWformer is the Sparse Wavelet Transformer Block (SWTB), which optimizes feature processing through two primary components:
Sparse Wavelet Self-Attention (SWSA)
Standard self-attention is computationally expensive and often processes irrelevant spatial noise. SWSA utilizes DWT to decompose features into low-frequency (LF) and high-frequency (HF) components. The HF branch employs ReLU activation to enforce sparsity, effectively filtering out noise, while the LF branch uses Softmax to preserve essential structural information. These components are adaptively fused to focus the model's attention on the most informative features.
Multi-scale Wavelet Feed-Forward Network (MWFN)
Replacing traditional feed-forward layers, the MWFN applies recursive DWT to analyze images across multiple resolutions (from $1/2$ to $1/16$ scale). This multi-scale approach captures both fine local details and broad global context simultaneously, avoiding the high parameter count associated with heavy convolutional layers.
Experimental Results
USWformer was evaluated on standard benchmarks including UIEB, LSUI, and real-world datasets like U45, UCCS, and SQUID.
Quantitative and Qualitative Performance
USWformer achieves a PSNR of 25.68 dB on the UIEB dataset, surpassing previous state-of-the-art models like Spectroformer (24.96 dB). Visually, the model excels at restoring natural color balance and enhancing edge clarity in complex underwater scenes, such as coral reefs and shipwrecks.
Efficiency Analysis
A defining characteristic of USWformer is its extreme efficiency. With only 1.19M parameters and 9.14 GFLOPs, it is significantly smaller than competitors like WaterNet (24.8M parameters) and U-shape (65.6M parameters). This represents a nearly 55-fold reduction in size compared to U-shape while delivering superior image quality.
Conclusion
USWformer demonstrates that architectural intelligence and frequency-domain awareness are more critical than raw model capacity for underwater image enhancement. By utilizing sparse wavelet transforms, the model provides a high-performance, lightweight solution suitable for resource-constrained marine applications. Future work will explore the impact of different wavelet bases and the extension of this approach to high-resolution video streams.
Summary updated