RT-CBAM: Refined Transformer Combined with Convolutional Block Attention Module for Underwater Image Restoration

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Transformers show promise in visual tasks, but struggle with fine local features in complex underwater images.
• Conventional Convolutional Neural Networks (CNNs) are adept at local features but lack the global context of transformers.

Purpose of the Study:

• To develop an improved transformer model for enhanced underwater image processing.
• To address limitations in capturing both local and global features for better detail and color restoration.

Main Methods:

• Proposed a refined transformer model with improved feature blocks (dilated transformer block) for accurate attention computation.
• Integrated a self-supervised local and global blind-patch network in the bottleneck layer for enhanced detail and texture recovery.
• Introduced a multi-scale convolutional block attention module (MSCBAM) to improve color channel feature representation and restoration.

Main Results:

• The refined transformer combined with convolutional block attention module (RT-CBAM) demonstrated superior performance in detail processing and color restoration.
• Achieved best results compared to two traditional methods and six deep learning approaches.
• The model effectively captures both local and global features, crucial for complex underwater imagery.

Conclusions:

• The RT-CBAM model offers significant advancements in underwater image restoration.
• The proposed methods enhance the model's ability to process fine local details and restore color information.
• The model is suitable for deployment on underwater robot sensors for real-world ocean exploration.