Research About Digital Image Foreground Extraction

Digital image foreground extraction is a hot computer vision issue. Foreground extraction, also known as matting, is an important computer vision processing technology and the basis of many other image-processing algorithms. Matting is to divide image into two parts, foreground region and the background region and extract the foreground image in order to save it or blend it with another image. In practices, industry, medical and media field all relies heavily on the image matting to process and analyze image data?Research on matting algorithm has important social value and theoretical importance.The main work of this paper is to study single frame based image matting algorithm. Existing matting algorithms gets two problems: requires many manual processes and runs slowly. This paper propose two improvements on existing algorithms: algorithms automation and parallel acceleration. This paper consist of the following aspects:1) This paper introduce the goals and basic methods of matting algorithm,including the general process flow. We make conclusions of existingmatting algorithm and analyze the advantages and disadvantages of them.Based on the two commonly existing problems, we introduce our researchcontents of our paper. The article also introduces several classicalalgorithm evaluation methods and data sets.2) This paper proposed an automatic tongue image extraction method. Firstly,we train a tongue body detector to get initial location of tongue. Then weanalyze color information of the sample region and use iterative histogramprojection to detect tongue body location and create corresponding trimap.Three types of trimap are proposed, they can handle different types of inputimages. Finally, use matting algorithm to precisely calculate tongue regionand extract tongue body. Experimental results prove that the proposedmethod can handle tongue body extraction precisely and meet the practicalrequirements. 3) Learning based digital matting(LBDM) is precise but slow and cannot meetrequirements of real-time usage. This paper propose parallel acceleratedLBDM method. LBDM algorithm consist of two stages: learning stage andsolving stage. Proposed parallel algorithm accelerate LBDM using thecombined power of both GPU and CPU. Firstly, use GPU to calculate thetask of each pixel in the learning stage in parallel. Then, use CPU tocomplete the task of each image block in the solving stage in parallel. Wepropose trimap skeleton based blocking algorithm(TSB) to divide imageinto blocks. TSB generates better blocks than the traditional algorithms.Experimental results proves that proposed method achieves up to 12 timesof speedup under given hardware, which significantly reduces algorithmtime consumption.