Deep Interactive Image Matting

Abstract

Image Matting for Compositing is the cutting of an object from an image for background replacement. It is an interactive process fundamental to image editing. Useful matting algorithms both reduce the amount of interaction necessary and accurately compute transparency and foreground colours in difficult cases like translucency and thin structures like hair. In 2016, the first deep-learning based matting methods were proposed with significant accuracy improvements over previous techniques. Despite excellent accuracy on difficult images, they fail to work well without copious user input even in simple cases where classic algorithms succeed. Concurrently, the first deep-learning based interactive binary-matting(segmentation) methods were proposed and successful in reducing the amount of user input necessary for a rough selection. Although these are practical for some simple images, they are not capable of real matting i.e finely detailed non-binary selections. In this thesis, we develop a novel deep learning-based matting algorithm which combines the strengths of interactive segmentation and matting. Specifically, we propose `interactivity focused' training and architecture for deep interactive segmentation methods to allow for more finely detailed user edits and selections. For matting, we modify our segmentation architecture to allow for accurate joint alpha and foreground colour prediction. Finally, we propose a method to finetune our matting algorithm such that it retains its accuracy but also is capable of quality selections with far less user input. Our interactive segmentation, matting and interactive matting algorithms achieve state-of-the-art performance on various challenging benchmarks, demonstrating their effectiveness for practical applications.