수중 로봇을 위한 Vision Transformer 기반의 바이오파울링 이상탐지 (Vision Transformer-based Biofouling Anomaly Detection for Underwater Robots)

서지정보

· 발행기관 : 한국로봇학회
· 수록지 정보 : 로봇학회 논문지 / 20권 / 2호 / 310 ~ 321페이지
· 저자명 : 장현수, 이동규, 노세환, 강동완, 유선철, 김아영

초록

Detection and removal of biofouling is crucial for maintaining underwater structure, since biofouling can lead to various operational problems. Potential risk and inefficiency of human detection in underwater environment led to development in robotic systems for automated detection. Recent studies demonstrate the effectiveness of deep learning in automated biofouling detection, but such methods often rely on numerous amounts of labelled data, which is scarce and difficult to obtain due to underwater conditions such as attenuation and scattering. This paper proposes an anomaly detection model using a vision transformer (ViT) for biofouling detection in underwater images. Unlike convolutional neural network, ViT can better capture the global context of images, making it more suitable for textureless underwater images. We further address data scarcity by applying a pre- trained foundation model and finetuning it on underwater dataset. We show that our method outperforms CNN-based biofouling detection model by qualitative and quantitative evaluations. To show that our method is fit for real-time ocean environment, we collected biofouling dataset from floating structure and assessed our method with the dataset.

영어초록

Detection and removal of biofouling is crucial for maintaining underwater structure, since biofouling can lead to various operational problems. Potential risk and inefficiency of human detection in underwater environment led to development in robotic systems for automated detection. Recent studies demonstrate the effectiveness of deep learning in automated biofouling detection, but such methods often rely on numerous amounts of labelled data, which is scarce and difficult to obtain due to underwater conditions such as attenuation and scattering. This paper proposes an anomaly detection model using a vision transformer (ViT) for biofouling detection in underwater images. Unlike convolutional neural network, ViT can better capture the global context of images, making it more suitable for textureless underwater images. We further address data scarcity by applying a pre- trained foundation model and finetuning it on underwater dataset. We show that our method outperforms CNN-based biofouling detection model by qualitative and quantitative evaluations. To show that our method is fit for real-time ocean environment, we collected biofouling dataset from floating structure and assessed our method with the dataset.

참고자료

· 없음