Improving rare-class detection in deep-sea imagery via generative augmentation with stable diffusion.

Area of Science:

• Deep-sea ecology
• Computer vision
• Artificial intelligence

Background:

• Megabenthos are vital for deep-sea ecosystem stability, but their detection is hindered by data scarcity for rare species.
• High costs of deep-sea exploration limit data acquisition, exacerbating the challenge of identifying underrepresented species.

Purpose of the Study:

• To develop and evaluate a data augmentation framework to address data scarcity for rare deep-sea benthos.
• To improve the accuracy and efficiency of deep-sea megabenthic detection using advanced AI techniques.

Main Methods:

• A Stable Diffusion (SD) model was fine-tuned using Low-Rank Adaptation (LoRA) to synthesize rare benthos images.
• ControlNet was employed to integrate generated images into deep-sea backgrounds with controlled layouts and automatic bounding-box annotation.
• The framework was tested on datasets from optically tethered underwater vehicles (OTVs) and autonomous underwater vehicles (AUVs), focusing on 7 rare species.

Main Results:

• The synthesized images achieved competitive quality metrics (FID: 117.11, IS: 4.97).
• The data augmentation strategy significantly improved detection performance (AP50-95 and AP50) on both OTV and AUV datasets compared to baseline methods.
• Notable performance gains were observed for rare species, with improvements up to 23.6% on the OTV dataset and 15.1% on the AUV dataset.
• The proposed method outperformed traditional augmentation techniques.

Conclusions:

• The proposed data augmentation framework effectively addresses data scarcity for rare deep-sea benthos.
• This AI-driven approach enhances deep-sea megabenthic detection accuracy, offering a valuable tool for conservation.
• The method demonstrates superior performance over traditional augmentation, highlighting its utility in deep-sea surveys.