Researches On Video Foreground Automatic Extraction Algorithm Based On Eyes Detection

For image and video are the main forms of information, the editing research ofthem had been becoming the focus of study. Based on the eyes detection, we proposeda video foreground automatic extraction algorithm including the following three parts:Firstly, the definite foreground and definite background were obtained to complete theautomatic interactive process through the eyes detection; Secondly, the first image ofvideo utilized the definite area to get the foreground automatedly; At last, by thecontinuity assumption, we could expand the first foreground of image to the otherimages and get the foreground of video.1. Interactive algorithm based on eyes detection: This paper proposed animproved Real-Adaboost algorithm, and used the union of EOH (edge of histogram)feather and Haar feather insteading of the single feather to get a more effective androbust trainning and testing model for eye detection. The experiments showed theproposed algorithm could increase the detection rate and decrease the trainning time.Then some areas belonging to definite foreground and definite background could befixed, so this process could replace the traditional human interactive.2. Foreground extraction algorithm of image: In order to expand the definiteforeground and definite background into the whole picture, based on the KDE (kerneldensity estimation) probability of bilayer segmentation, this paper proposed theforeground extraction algorithm based on geodesic distance. Fistly, through thebilayer segmentation and get the trimap, then we utilized the geodesic distance for thefuture foreground extraction, at last the foreground was captured more accurately andrapidly.3. Foreground extraction algorithm of video: Based on image foregroundextraction of the first image, according to the continuity assumption and KDEprobability, we proposed a new method for caculating the probability of the pixelbelonging to foreground and backgrond by the integral image. Then the trimap of thecurrent image could be constructed; next the geodesic distance was used for theextraction of small edge; finally, by some morphological technology, the videoforeground could be extracted fast, robustly and effectively.In the end, we summarized the work of this paper and illustrated the deficiencyof this paper. Also the further research was pointed out.