Automated detection of whales in aerial imagery

点击翻译按钮获取中文摘要

Remote sensing, or capturing and processing images of the earth's surface from satellite or aerial platforms, can efficiently allow scientific monitoring of vast earth regions. As an example, remote sensing is essential for monitoring whale populations. Whales are biologically and environmentally important to Canada and part of food security as well as culturally and economically important for Inuit. Although many images can be captured quickly, a year is required to manually process the 1000s of images aerially captured in a single whale field survey. Our research endeavors to use automated computer vision algorithms to expedite this process. This task is challenging due to infrequent observation of whales, the small size of the whale relative to the image size, and the often unclear whale appearance due to the unpredictable ocean appearance. Novel machine learning models for computer-aided processing and analysis of the large-volume complex images can be used to perform automated whale population counts, and thereby process survey image data more efficiently and consistently. Advancements will be made with respect to improved capture of visual cues to identify what could be a whale and context-aware methods using prior knowledge to distinguish whales from look-alikes. These new methods will facilitate the development of remote sensing software and analytic solutions for whale identification in field survey images. The methods can then also support other remote sensing target detection and classification tasks in ocean monitoring applications.