[Paper Review] Jointly Adversarial Network to Wavelength Compensation and Dehazing of Underwater Images

Xueyan Ding, Yafei Wang+4 more|arXiv (Cornell University)|Jul 12, 2019

Image Enhancement Techniques43 references15 citations

TL;DR

This paper proposes a jointly adversarial network (JWCDN) that simultaneously performs wavelength compensation and dehazing in underwater images by embedding a simplified underwater image formation model into a generative adversarial network. The method jointly estimates transmission map, wavelength attenuation, and background light using a multi-scale densely connected encoder-decoder with gradient-preserving loss, achieving state-of-the-art results on synthetic and real-world datasets.

ABSTRACT

Severe color casts, low contrast and blurriness of underwater images caused by light absorption and scattering result in a difficult task for exploring underwater environments. Different from most of previous underwater image enhancement methods that compute light attenuation along object-camera path through hazy image formation model, we propose a novel jointly wavelength compensation and dehazing network (JWCDN) that takes into account the wavelength attenuation along surface-object path and the scattering along object-camera path simultaneously. By embedding a simplified underwater formation model into generative adversarial network, we can jointly estimates the transmission map, wavelength attenuation and background light via different network modules, and uses the simplified underwater image formation model to recover degraded underwater images. Especially, a multi-scale densely connected encoder-decoder network is proposed to leverage features from multiple layers for estimating the transmission map. To further improve the recovered image, we use an edge preserving network module to enhance the detail of the recovered image. Moreover, to train the proposed network, we propose a novel underwater image synthesis method that generates underwater images with inherent optical properties of different water types. The synthesis method can simulate the color, contrast and blurriness appearance of real-world underwater environments simultaneously. Extensive experiments on synthetic and real-world underwater images demonstrate that the proposed method yields comparable or better results on both subjective and objective assessments, compared with several state-of-the-art methods.

Motivation & Objective

• Address the dual challenges of color distortion and reduced visibility in underwater images caused by light absorption and scattering.
• Overcome limitations of prior methods that model only scattering or absorption separately, by jointly modeling both physical processes.
• Develop a novel underwater image synthesis method based on inherent optical properties to simulate realistic degradation across different water types.
• Improve image quality by preserving edges and textures through a gradient-preserving loss and edge-enhancing module.
• Achieve superior performance in both subjective and objective evaluations compared to state-of-the-art methods.

Proposed method

• Embed a simplified underwater image formation model into a GAN framework to jointly estimate transmission map, wavelength attenuation, and background light via dedicated network modules.
• Design a multi-scale densely connected encoder-decoder network with a gradient-preserving loss to enhance transmission map estimation and reduce halo artifacts.
• Introduce an edge-preserving network module to refine fine details and restore sharpness in recovered images.
• Propose a novel underwater image synthesis method using inherent optical properties (absorption and scattering coefficients) to simulate realistic color casts, contrast degradation, and blurriness across diverse water types.
• Train the network using a combination of perceptual, adversarial, and content-preserving losses to improve realism and structural fidelity.
• Utilize math operation modules to directly integrate the physical image formation model into the GAN architecture for end-to-end optimization.

Experimental results

Research questions

• RQ1 Can a unified deep learning framework jointly optimize wavelength compensation and dehazing in underwater images more effectively than separate processing?
• RQ2 How does modeling both surface-object path attenuation and object-camera path scattering improve image restoration compared to modeling only one process?
• RQ3 To what extent can a GAN-based approach with physical priors outperform existing state-of-the-art methods in both quantitative and qualitative assessments?
• RQ4 Can a synthesis method based on inherent optical properties generate realistic underwater images that better train and evaluate enhancement networks?
• RQ5 What is the contribution of the multi-scale densely connected encoder-decoder and edge-preserving module to transmission map estimation and final image quality?

Key findings

• The proposed JWCDN achieved the lowest blur metric (0.21473) and highest UIQM (1.5372) on the RealB dataset, outperforming all compared methods.
• On the RealA dataset, JWCDN achieved a blur metric of 0.21818 and UIQM of 1.574, demonstrating superior performance in both objective and perceptual quality.
• The ablation study showed that removing the content-preserving loss (JWCDN-woCL) increased blur, while removing the edge-preserving module (JWCDN-woEP) caused color distortion and over-bright regions.
• Quantitative results on the SyntTest dataset showed that JWCDN achieved a PSNR of 21.096, SSIM of 0.75712, and PCQI of 0.76362, significantly outperforming variants without key modules.
• The proposed underwater image synthesis method successfully simulated realistic degradation patterns across different water types, enabling more robust training and evaluation.
• Visual comparisons confirmed that JWCDN produced sharper edges, better color restoration, and reduced halo effects compared to Retinex-based, GDCP, UDCP, IBLA, and UGAN methods.