IA2U: A Transfer Plugin with Multi-Prior for In-Air Model to Underwater

TL;DR

This work tackles the difficulty of applying in-air vision models to underwater scenes by introducing IA2U, a plug-and-play transfer plugin that embeds three underwater priors—water type, degradation, and sample—into a Transformer-like feature refinement framework. IA2U comprises a dual-branch architecture with a Feature Enhancement Network and an Underwater Prior Generator, and employs a Multi-Scale Feature Aggregation module to perform robust, hierarchical enhancement guided by IE and OD losses. The method generates a Prior P from priors and intermediate features to steer attention and feature fusion, enabling improved underwater image enhancement and object detection when paired with pre-trained in-air models. Extensive experiments on UIE and UOD demonstrate consistent gains across multiple detectors and UIE baselines, with ablations confirming the value of each prior and the full-scale alignment strategy. The authors also commit to releasing the code, underscoring IA2U’s practical potential for broad deployment in underwater vision tasks.

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.

Figures (8)

• Figure 1: Illustrate the overall process of (a) image enhancement, (b) task-driven image enhancement, (c) feature enhancement, and (d) proposed IA2U. In particular, the IE for image enhancement. IFE presents image feature enhancement. OD for object detection, and IE/OD Loss is the loss function for the corresponding task.
• Figure 2: Overview architecture of the IA2U method. The underwater image is passed through the classifier to obtain two prior knowledge of water type and degradation features and combined with the features in FEN to generate Query. Next, the generated Query is fed to the Attention module in FEN for feature refinement. Finally, the feature increments are combined with the original image through the residual connection to feed the subsequent IE and OD network.
• Figure 3: Prior Selection in IA2U Model. Top row: Degradation prior knowledge in various water types, emphasizing different degradation levels. Bottom row: Sample prior knowledge, categorized by differences in ambient lighting and object characteristics.
• Figure 4: Structure of the Priori Query Generator (PQG). Depicting classifier predictions (CLASS) and middle layer features (R), with $F_{j}$ representing the output of the $j$th layer of the FEN network
• Figure 5: Structure of Multi-Scale Feature Aggregation (MSFA): Integrating Depth-Wise Convolution (DWC) and Adaptive Average Pooling (AAP).