Adaptive Pixel Pair Evaluation Method for Image Matting

Image matting is an ill-posed problem that aims to extract the opacity of foreground objects in an image. Pixel-pair-optimization-based (PPO-based) image matting approaches are widely adopted in natural image matting, whereby the alpha value is estimated by choosing the optimal pixel pair according to a pixel pair evaluation (PPE) function. Multiple PPE criteria are employed to improve the accuracy of PPE, resulting in the weight setting problem of PPE criteria. Existing PPE functions use fixed weight PPE criteria, which cannot provide the accuracy of PPE on the little transparent images due to the satisfaction degree of PPE criteria related to the type of the image. To address this shortcoming, in this work, an adaptive weight criteria PPE method is presented, which adaptively adjusts the contribution of chromatic distortion and spatial closeness criteria to the PPE function by analyzing the type of the image. Experimental results show that the proposed adaptive weight criteria PPE method provides accurate PPE compared with existing PPE methods, especially on the little transparent Images.