Digital Image Matting: A Review

Introduction

Digital Image Matting is a way through which we can extract foreground object from the given image. Matting problem means the problem of extracting the foreground accurately, so that the extracted foreground when composites with new backgrounds, should appear realistic. The new composite image should appear as real as if actually taken by the camera itself.

Image Matting is different from segmentation technique. In segmentation technique, each pixel either completely belongs to the background or completely belongs to the foreground, but in Image Matting some pixels may belong to foreground as well as background called partial pixels. Image Matting is all about calculating, how much fractions of foreground and background colors are mixed together to generate partial pixel color. This fraction of foreground color is called opacity/blending proportion/α value of the partial pixel.

Partial pixel is the pixel which neither completely belongs to foreground nor completely belongs to background region i.e. it receives color from both foreground and background regions. Generally, partial pixels are pixels along the boundary of an object or pixels that include fine objects like hair and fur.

For matting to be performed, first we need to divide the input image into three regions: Definite Foreground, Definite Background and Unknown region. This information is given by the user in the form of trimap. There are various user interface tools like soft scissor, intelligent scissor and component picking tools available, through which user can specify the trimap.

While specifying the trimap, the region for which user is sure that pixels completely belong to the background is specified as Definite Background region and the region for which user is sure that pixels completely belong to foreground is specified as a Definite Foreground region. The region for which user is not sure that whether pixels belong to the foreground or background, is specified as Unknown region. Ideally the unknown region should include only partial pixels, but it is not possible by the user to specify such an ideal trimap. Actually the unknown region includes partial pixels as well as some pure pixels (near the partial pixels).

The solution of the matting problem involves solving the mating equation for each pixel lying in the unknown region called an unknown pixel (including partial pixel and some known foreground and known background pixels). Matting equation is given by Equation (1)..(1) where i represents pixel location in the image.

Here the is defined as the blending proportion of foreground color at each pixel (Porter & Duff, 1984). For definite foreground region: and and for definite background region: and. So, for definite foreground and definite background region value of F, B and α are known. We need to calculate them only for unknown pixels. Once we know the value α, F and B, we can replace B with B' (new background) in the matting Equation (1) and get the new output image having same foreground but changed background.

Wang and Cohen (2007a) categorized the matting techniques into three categories: sampling based, affinity based and optimization based techniques. In this paper, we are categorizing the matting techniques from the user’s input point of view. We categorize the matting techniques into two categories: Trimap based and Scribble based matting techniques.

The rest of the paper has been organized as follows: section 2 presents literature review, in which various matting approaches have been discussed along with their strengths and limitations. Section 3 gives a comparative analysis among the described matting approaches and finally section 4 concludes the paper.