Abstract
In underwater environments, variations in suspended particle concentration and turbidity cause severe image degradation, posing significant challenges to image enhancement (IE) and object detection (OD) tasks. Currently, in-air image enhancement and detection methods have made notable progress, but their application in underwater conditions is limited due to the complexity and variability of these environments. Fine-tuning in-air models saves high overhead and has more optional reference work than building an underwater model from scratch. To address these issues, we design a transfer plugin with multiple priors for converting in-air models to underwater applications, named IA2U. IA2U enables efficient application in underwater scenarios, thereby improving performance in Underwater IE and OD. IA2U integrates three types of underwater priors: the water type prior that characterizes the degree of image degradation, such as color and visibility; the degradation prior, focusing on differences in details and textures; and the sample prior, considering the environmental conditions at the time of capture and the characteristics of the photographed object. Utilizing a Transformer-like structure, IA2U employs these priors as query conditions and a joint task loss function to achieve hierarchical enhancement of task-level underwater image features, therefore considering the requirements of two different tasks, IE and OD. Experimental results show that IA2U combined with an in-air model can achieve superior performance in underwater image enhancement and object detection tasks. The code will be made publicly available.
Introduction
Underwater computer vision is notoriously difficult due to light attenuation, scattering, and suspended particles. These factors cause severe color shifts, low contrast, and "hazy" turbidity that vary wildly depending on the specific body of water. While "in-air" models trained on terrestrial datasets like COCO or ImageNet are highly advanced, they perform poorly underwater. Researchers face a dilemma: building specialized underwater models from scratch is data-intensive and expensive, but simply fine-tuning in-air models often fails to capture the unique physics of underwater degradation. Furthermore, there is often a conflict between Image Enhancement (IE)—making images look good for humans—and Object Detection (OD)—extracting features for machines. IA2U (In-Air to Underwater) is a "plug-and-play" transfer plugin designed to bridge this gap by adapting standard terrestrial models for harsh underwater conditions.
The IA2U Architecture
The IA2U framework acts as a specialized pre-processor, "translating" underwater features into a format the in-air model can better understand through a dual-branch architecture.
- Underwater Prior Generation Network (PGN): This branch uses a pre-trained, frozen classifier to categorize the input image into one of nine water types. It extracts three "weaker" priors: the Water Type Prior (general color and visibility), the Degradation Prior (microscopic differences in detail and texture), and the Sample Prior (specific environmental conditions at the time of capture).
- Feature Enhancement Network (FEN): This branch processes image features using a novel Multi-Scale Feature Aggregation (MSFA) module. Unlike traditional methods that pick a single "anchor" scale, MSFA uses a "full-scale" strategy to align features across different receptive fields (1x1, 3x3, and 5x5 kernels) to minimize information loss during resampling.
- Priori Aggregation: Inspired by Transformers, the plugin uses the environmental priors from the PGN as a "Query" to interact with the image features (Key/Value). This allows the model to adaptively filter and enhance features based on the specific type of water degradation present.
Experimental Evaluation
The IA2U plugin was evaluated using several key underwater benchmarks, including UIEBD, LSUI, and RUOD. The results demonstrate that the plugin significantly boosts performance in both underwater image enhancement and object detection without requiring a total model redesign.
- Object Detection: Adding IA2U to a standard Faster R-CNN improved Average Precision (AP) on the RUOD dataset by 3.2%. It showed similar gains across other detectors like RetinaNet and TOOD.
- Image Enhancement: On the UIEBD dataset, IA2U boosted the PSNR of the "Shallow-UWNet" baseline by 5.23 dB and the SSIM by 0.06.
- Qualitative Superiority: Visual results show that IA2U-equipped models are significantly better at removing the "greenish" or "bluish" tints common in deep-sea photography compared to standalone dehazing or enhancement models.
Conclusion
IA2U effectively bridges the domain gap between in-air and underwater environments by treating environmental knowledge as a query in an attention mechanism. This "un-lost" information strategy ensures that features from all scales are preserved and refined. While the system currently relies on a frozen classifier, future work may explore more diverse priors or integration into more complex backbones to further improve performance on real-time underwater platforms.
Summary updated