A diffusion model-based image generation framework for underwater object detection

Area of Science:

• Computer Vision
• Marine Technology
• Data Augmentation

Background:

• Underwater object detection is vital for marine monitoring and rescue.
• Limited underwater data and scene diversity impede detection accuracy.
• Existing methods often rely on external data, creating domain gaps.

Purpose of the Study:

• To propose a novel framework, Underwater Layout-Guided Diffusion Framework (ULGF), for augmenting underwater object detection datasets.
• To generate high-fidelity and diverse underwater images using only a small set of existing underwater data.
• To improve the performance of underwater object detection models.

Main Methods:

• Developed the Underwater Layout-Guided Diffusion Framework (ULGF), a diffusion model.
• The framework operates exclusively on limited underwater images and their labels.
• No external or in-air data integration is used for image generation.

Main Results:

• ULGF generates theoretically infinite high-fidelity and diverse underwater images.
• The generated images exhibit a smaller domain gap compared to existing methods.
• Substantially enhanced object detection performance in real-world underwater scenarios.

Conclusions:

• The ULGF framework effectively addresses the challenge of limited underwater data.
• It provides a novel approach for synthetic data generation in underwater imaging.
• The method shows significant potential for advancing marine technology and applications.