Interactive Matting Technology On The Image And Image Sequence

Matting is a technology of extracting visual objects from given images, and is widely used for preparing materials for image composting at the post-processing stage. The foreground and background are separated by estimating a mask image, which represents how the foreground image is mixed with the background image. If only 0 and 1 is used in the mask image, we call it 0-1 matting; otherwise we call it alpha matting. In the thesis, GrabCut based 0-1 matting is implemented. To speed up matting on large images, the following improvements are applied to GrabCut:down-sampling and pre-segmentation. The first improvement performs small scale Grabcut segmentation several times on the down-sampled images instead of one large scale Grabcut on original image, and improves matting speed significantly without losing accuracy. The second improvement builds graph on over-segmented images blocks, and turns a large scale problem into small scale problems. Experiment results show that both the modifications can improve processing speed significantly. Alpha matting is commonly used to separate transparent objects and small objects, such as glasses and hair, from their background. In this thesis, Robust Matting based on alpha estimating is used for extracting such visual objects, and automatic trimap generation based on 0-1 matting is researched. The experiments show that matting results using auto-generated trimap and manual trimap are very close, but automatic trimap generation saves lots of human interaction.Although image sequences can be processed in a sequential manner, it requires too much human interaction and too much time consuming. In this thesis, we build graph on multi images based on their spatio-temporal relationships, and cut the images within a single round of Grabcut. A significant advantage of spatio-temporal Grabcut is the complexity of human interaction is reduced from n times to only one time, which means to extract foreground images from an image sequence of size n, we only need to interact on one of the images, instead of n images. To speed up spatio-temporal Grabcut and make the process more responsive, we propose applying spatio-temporal Grabcut on a k+1 sliding window instead of the entire image sequence. Spatio-temporal Grabcut gives very reasonable segmentation result on both simple and complex image sequences.