ViT-UWA: Vision Transformer Underwater-Adapter for Dense Predictions Beneath the Water Surface.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Vision Transformers (ViT) excel in computer vision but struggle with underwater image challenges.
• Underwater images present difficulties like scattering, degradation, and complex environments, hindering dense prediction task performance.

Purpose of the Study:

• To develop a novel Vision Transformer Underwater-Adapter (ViT-UWA) for improved underwater dense prediction.
• To create a detail-focused ViT backbone adaptable for various underwater computer vision tasks without task-specific pretraining.

Main Methods:

• Introduced High-frequency Components Prior (HFCP) to restore lost high-frequency details in underwater images.
• Developed a Detail Aware Module (DAM) for multi-scale feature extraction, creating a detail-focused feature pyramid.
• Implemented ViT-DAM Cross Fusion (VDCF) for bidirectional feature fusion between ViT and DAM.

Main Results:

• ViT-UWA demonstrated state-of-the-art performance on underwater semantic segmentation, instance segmentation, and object detection.
• ViT-UWA-B achieved 46.4 box AP and 44.2 mask AP on the USIS10K dataset with ImageNet-22K pretraining.

Conclusions:

• The proposed ViT-UWA effectively addresses the challenges of underwater dense prediction tasks.
• ViT-UWA offers a superior, detail-focused approach for underwater computer vision applications.