DS-SIAUG: A Self-Training Approach Using a Disrupted Student Model for Enhanced Side-Scan Sonar Image Augmentation

Area of Science:

• Marine technology
• Artificial intelligence
• Image processing

Background:

• Accurate subsea target detection using side-scan sonar is crucial for seabed analysis.
• The accuracy of intelligent target recognition heavily relies on the quantity and representativeness of sonar image samples.

Purpose of the Study:

• To develop a novel data augmentation method for side-scan sonar images to enhance intelligent target recognition.
• To improve the accuracy and efficiency of subsea target detection through advanced image processing techniques.

Main Methods:

• A novel augmentation method, Disrupted Student Self-training Augmentation (DS-SIAUG), was designed.
• The method utilizes an adversarial network combining Denoising Diffusion Probabilistic Model (DDPM) and You Only Look Once (YOLO) for image augmentation.
• A Disrupted Student model was employed to filter and select representative target images for iterative retraining.

Main Results:

• The DS-SIAUG method achieved target recognition accuracy comparable to manual selection.
• Intelligent target recognition accuracy was improved by approximately 5% compared to direct adversarial network augmentation.
• The Disrupted Student model effectively filters representative images, enhancing the training dataset's quality.

Conclusions:

• The proposed DS-SIAUG method offers an effective approach to augment side-scan sonar image datasets.
• This technique significantly improves the accuracy of intelligent subsea target recognition.
• DS-SIAUG provides a valuable tool for enhancing marine technology applications reliant on sonar data analysis.