Unsupervised segmentation of low depth of field images based on L0 regularized matting model

An effective unsupervised segmentation method is proposed to extract the object of interest from low depth of field images based on a novel L0 regularized matting model. First of all, a multi-scale reblurring model is utilized, together with guided filter and morphological filter, to generate a trimap that roughly labels the focused and defocused regions. Then, an L0 regularized matting model is proposed to obtain the accurate segmentation of the object of interest. Experimental results demonstrate that the proposed method achieves state-of-the-art performance for unsupervised segmentation under various situations, and is robust to noise.