Improved Bayesian alpha matting in YIQ color space

Research focuses on an improved Bayesian matting algorithm combining the segment and digital matting approaches in YIQ color space. A simplified intelligent scissors is used to segment the image to three regions: foreground, background and unknown regions. An improved Bayesian matting method is proposed by modeling distribution of alpha value as a Gaussian distribution and introduced image gradient to weight the standard deviation of alpha distribution. This improvement makes the alpha estimation more accurate. Meanwhile, alpha value is estimated in YIQ color space instead of in RGB color space. Color is separated into chroma and intensity information. The importance is emphasized to avoid errors brought by their interaction. Experimental results demonstrate that proposed algorithm generates better matte than existing matting techniques on many complex natural images.