The advances in Autonomous Underwater Vehicle (AUV) and photographic technology have enabled us to collect large volumes of visual data capturing vast areas of the seafloor for monitoring marine ecosystems and environments. The problem we are now facing is development of efficient and effective methods of analysing these data in reasonable time frames. Detection and classification of Objects of Interest (OOI) such as benthic fauna is usually performed by experts. This task can be extremely time consuming and in many cases infeasible. Supervised machine learning approaches have received a great deal of attention in recent years which has partly been due to the advances in computing power but also due to the availability of large and accurately annotated datasets. In this talk I will detail why annotation of large underwater image datasets is a challenging problem, explain what is meant by computer vision and demonstrate how techniques from this field can be used to enable faster acquisition of ecological information compared to relying strictly on experts alone. This will include presentation of results produced by applying Convolutional Neural Networks (CNN) to learn the visual characteristics of deepsea fauna found in the Porcupine Abyssal Plain (PAP) - these deep neural networks allow us automatically identify a species from just an image. Finally I will discuss plans for future work, including how we might use computer vision to better understand areas of the seafloor without any prior expert knowledge.
Thursday 19 October 2017 - 14:00 to 15:00
NOC Southampton - Node Room (074/02) (Waterfront Campus).
Dr William Hosking (NOCS)