Abstract
Underwater images suffer from serious color distortion and detail loss because of the wavelength-dependent light absorption and scattering, which seriously influences the subsequent underwater object detection and recognition. The latest methods for underwater image enhancement are based on deep models, which focus on finding a mapping function from the underwater image subspace to a ground-truth image subspace. They neglect the diversity of underwater conditions which leads to different background colors of underwater images. In this paper, we propose a Class-condition Attention Generative Adversarial Network (CA-GAN) to enhance an underwater image. We build an underwater image dataset which contains ten categories generated by the simulator with different water attenuation coefficient and depth. Relying on the underwater image classes, CA-GAN creates a many-to-one mapping function for an underwater image. Moreover, in order to generate the realistic image, attention mechanism is utilized. In the channel attention block, the feature maps in the front-end layers and the back-end layers are fused along channels, and in the spatial attention block, feature maps are pixel-wise fused. Extensive experiments are conducted on synthetic and real underwater images. The experimental results demonstrate that CA-GAN can effectively recover color and detail of various scenes of underwater images and is superior to the state-of-the-art methods.
Introduction
Capturing high-quality images underwater is a significant challenge. Unlike terrestrial environments, water absorbs light at varying rates depending on the wavelength—red light is absorbed most rapidly, followed by yellow and green, often leaving images with a pervasive blue or green tint. Additionally, suspended particles cause light scattering, resulting in a "foggy" appearance and a substantial loss of fine detail.
A major limitation of existing deep learning models is their attempt to find a single "one-size-fits-all" mapping from degraded images to clear ones. However, underwater conditions are highly diverse; shallow tropical waters differ significantly from turbid coastal environments. Ignoring these variations often leads to poor generalization and artifacts. The CA-GAN framework addresses this by treating enhancement as a conditional task, leveraging water type classification to guide the restoration process.
Proposed Methodology
Overall Architecture
The CA-GAN architecture is designed to handle the diversity of underwater environments through a many-to-one mapping approach. The generator incorporates a Water Class Embedding Block (WCEB) and a novel attention-based fusion mechanism.
As shown in the architecture, the model first classifies the input image into one of ten "water types" based on depth and turbidity. This class information is converted into a one-hot vector and processed through a Fully Connected (FC) layer to influence the WCEB. This allows the network to "normalize" the feature space based on the specific degradation characteristics of that water type.
Attention-Based Feature Fusion
To preserve fine details and ensure realistic textures, the authors introduce the Concurrent Channel and Spatial Attention Feature Fusion (CS-AFFB) module. This module recalibrates and fuses feature maps from both the front-end (encoder) and back-end (decoder) of the network.
The CS-AFFB operates through two primary paths:
- Channel Attention: Reweights feature channels to emphasize critical information, such as specific color components that may have been heavily attenuated.
- Spatial Attention: Focuses on pixel-wise details to ensure sharp edges and clear structural integrity.
By fusing these features, the model effectively "remembers" high-frequency details that are often lost during the standard encoding-decoding process of a GAN.
Experimental Results
The researchers evaluated CA-GAN using the new RSUIE (Realistic Single Underwater Image Enhancement) dataset, which includes 70,000 synthetic image pairs across 10 Jerlov water types, as well as real-world datasets like UIQS and UCCS.
Quantitative and Qualitative Performance
CA-GAN consistently outperformed state-of-the-art methods such as Water-Net and DPATN. On the RSUIE test set, it achieved superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) scores. For instance, in highly turbid "Type-9" water, CA-GAN reached a PSNR of 22.92, compared to 18.64 for Water-Net. Visually, the results show more natural color balance—particularly in the red channels—and sharper details without the over-saturation or halo artifacts common in other approaches.
Conclusion
CA-GAN represents a significant advancement in underwater image enhancement by explicitly modeling the diversity of water conditions. By combining class-conditioned normalization with a dual-attention fusion mechanism, the framework effectively restores both color and detail across a wide range of underwater scenes. Future work will focus on reducing the dependency on the initial classifier and optimizing the model for real-time deployment on low-power underwater platforms.
Summary updated