Categorising videos using a personalised category catalogue

Video is an extremely effective way of reaching farmers with the latest agricultural technology and stories of other farmers. With a well-organised multifaceted video library, we can provide the farmers with services such as easy navigation, search and recommendations of videos as per their needs. Since categories and tags assigned by video uploaders on YouTube are often prone to noise and are not uniform across a video collection, we propose a semi-automated system to achieve this desired video categorisation. We adopt active learning as our strategy to evolve a personalised category catalogue for agricultural videos. We present a multi-label classification system, with a large global category catalogue, such as Wikipedia, as its initial label space. We narrow down on a domain-specific category catalogue, evolving our system by using associative Markov networks with the categories as nodes. Each node incorporates structural constructs and a binary SVM classifier as node features. We show that our categorisation for agricultural videos is less granular and more uniform as compared to YouTube tags, while staying sufficiently specific.