[Paper Review] A Fusion Adversarial Underwater Image Enhancement Network with a Public Test Dataset

Hanyu Li, Jingjing Li+1 more|arXiv (Cornell University)|Jun 17, 2019

Image Enhancement Techniques42 references124 citations

TL;DR

The paper introduces a fusion adversarial network (FGAN) for underwater image enhancement and a public test dataset (U45), achieving competitive quality with fast testing and fewer parameters, along with ablation studies and application tests.

ABSTRACT

Underwater image enhancement algorithms have attracted much attention in underwater vision task. However, these algorithms are mainly evaluated on different data sets and different metrics. In this paper, we set up an effective and pubic underwater test dataset named U45 including the color casts, low contrast and haze-like effects of underwater degradation and propose a fusion adversarial network for enhancing underwater images. Meanwhile, the well-designed the adversarial loss including Lgt loss and Lfe loss is presented to focus on image features of ground truth, and image features of the image enhanced by fusion enhance method, respectively. The proposed network corrects color casts effectively and owns faster testing time with fewer parameters. Experiment results on U45 dataset demonstrate that the proposed method achieves better or comparable performance than the other state-of-the-art methods in terms of qualitative and quantitative evaluations. Moreover, an ablation study demonstrates the contributions of each component, and the application test further shows the effectiveness of the enhanced images.

Motivation & Objective

• Motivate the need for a public, representative underwater image test dataset capturing color casts, low contrast, and haze-like degradation.
• Propose a fusion adversarial network (FGAN) that blends two inputs to enhance underwater images with efficiency and accuracy.
• Design a multi-term loss (L_gt and L_fe) and spectral normalization to preserve content and improve image quality.
• Evaluate on the U45 dataset against several state-of-the-art methods and perform ablations to isolate component contributions.

Proposed method

• Use a two-input fully convolutional generator that fuses the raw underwater image with an image enhanced by a fusion-enhance method.
• Adopt an inception-inspired basic block with multi-scale convolutions and residual connections to improve feature learning.
• Employ a five-layer PatchGAN-style discriminator with spectral normalization (SN) for stable adversarial training.
• Define RaSGAN-based objective (L_D^RaSGAN, L_G^RaSGAN) combined with two feature-preserving losses L_gt and L_fe (L_gt(G) = E[||x − G(y)||1], L_fe(G) = E[||x_fe − G(y)||1]).
• Tune loss weights (λ_gt, λ_fe) to balance content preservation and fusion-enhanced feature fidelity.
• Evaluate on 256x256x3 inputs and compare against FE, RB, UDCP, UIBLA, DPATN, CycleGAN, WSCT, and UGAN.

Experimental results

Research questions

• RQ1 Can a fusion-based GAN architecture effectively correct underwater color casts while preserving scene details across diverse underwater conditions?
• RQ2 Does incorporating dual inputs and specialized loss terms improve objective and subjective underwater image quality compared to existing methods?
• RQ3 What is the impact of spectral normalization and architectural blocks on generation quality and testing efficiency?
• RQ4 How does the proposed method perform on a public, diverse underwater test dataset (U45) relative to state-of-the-art methods?
• RQ5 Do enhancements translate into benefits for downstream tasks (edge detection, object detection) in underwater scenes?

Key findings

• FGAN with dual-input fusion achieves competitive or superior UIQM and UCIQE scores across green, blue, and haze-like categories on U45 compared to several baselines.
• FGAN provides fast testing speed (0.0286 s per 256x256 image) with relatively few parameters, outperforming or matching larger models in speed.
• Ablation studies show that both spectral normalization and the proposed basic blocks contribute to performance gains, and increasing λ_fe improves alignment with fusion-enhanced inputs.
• On objective metrics, DPATN improves some scores but can yield visually poorer results in certain scenes; FGAn maintains stable high UIQM across subsets.
• Qualitative results indicate FGAN better corrects color casts and preserves details (e.g., coral in green scenes, fish in blue scenes) than several competing methods.
• Application tests reveal improved performance in downstream tasks like canny edge detection and object detection when using FGAN-enhanced images.