Addressing Computer Vision Challenges using an Active Learning Framework

Abstract

Machine Learning applications has transformed everyday life as well as industry by providing new successful opportunities in healthcare, transportation, banking, security, media monitoring and more. Computer Vision is an application of Machine Learning that recently has done a lot of progress, particularly in Face Recognition and Object Detection systems. These systems require large data sets to be trained with. Nevertheless, the available data sets contain large amounts of unlabelled samples. Active Learning is an innovative field that addresses the challenge of labelling large sets of unlabelled samples by leveraging only a small amount of manually labelled data. An efficient way of labelling a small amount of training data is utilizing user-friendly annotation tools. The latter allow playing a whole video streaming and capturing the desired entities. This interactive method could be very efficient as well as time-saving in comparison to traditional data collection methods. This thesis builds on state-of-the-art Face Recognition and Object Detection models, by implementing optimization methods that enhance the recognition accuracy. Further training is being introduced by making use of a robust Active Learning framework that results in creating extended data sets. Finally, our thesis proposes an integrated system, which involves effective techniques of associating face and object identification informa- tion, in order to extract as much knowledge as possible from a video streaming, in real-time.