Abstract: GrabCut (a foreground subject segmentation algorithm) is studied. The smoothing constraint of GrabCut is based on the two-point pixels of the image, that is, each pixel pair is regarded as a unit. However, the smoothness is point by point in essence. Therefore, based on the GrabCut, an image foreground subject segmentation algorithm (SAD_GrabCut for short, a GrabCut combining point by point and two-point pixel smoothing) is proposed, which combines the smoothing of point by point and two-point pixels and both of them are important. This importance could be two-point pixel constraint information or local density of point by point pixels. The empirical results on the images with relatively complex background environment, great similarity between the front and background, or the subject with sawtooth like edges (from COCO_test2014、DIV2K、BSDS300) show that SAD_GrabCut has certain competitiveness compared with GrabCut.