Underwater motion scene image restoration based on an improved U-Net network

Area of Science:

• Computer Vision
• Optical Engineering
• Image Processing

Background:

• Active underwater polarization imaging suppresses scattered light for improved clarity.
• Current methods require multiple images, limiting application to static scenes.
• Restoring underwater motion scenes remains a challenge due to image degradation.

Purpose of the Study:

• To develop a deep learning model for underwater image restoration from a single polarized image.
• To address the limitations of multi-image acquisition methods for dynamic scenes.
• To enhance the detail and quality of underwater images, particularly in motion.

Main Methods:

• Proposed a U-AD-Net deep learning model based on the U-Net architecture.
• Integrated Dense-Net and spatial attention modules to enhance feature extraction.
• Utilized polarization information from a single image as input for the network.

Main Results:

• The U-AD-Net model demonstrated superior performance in restoring underwater images.
• Restored images exhibited richer detail information compared to existing methods.
• Achieved dynamic adaptability for underwater moving scene restoration.

Conclusions:

• The proposed single-polarized image restoration method offers significant advantages for underwater imaging.
• U-AD-Net effectively extracts relevant polarization information for comprehensive scene restoration.
• This approach is highly suitable for real-time applications involving underwater motion scenes.