Automatic segmentation of focused objects from images with low depth of field

In this paper, we propose an automatic segmentation approach to extract focused objects from images with low depth of field (DOF). A focus energy map is first estimated based on the difference of high-frequency components between focused region and defocused background, and is exploited to construct region/boundary saliency maps on the basis of a pre-segmentation result by watershed transform. Then region/boundary masks for focused object are generated by entropy thresholding and flood filling, and an efficient boundary linking method is proposed to obtain closed region/boundary masks, which are exploited to reasonably generate a trimap containing seed regions for focused object and defocused background, and uncertain regions, respectively. Finally, the trimap is used as the input to an image matting model, which is utilized to classify the pixels in the uncertain regions to obtain an accurate focused object segmentation result based on the estimated alpha matte. Experimental results for a variety of low DOF images demonstrate the good segmentation performance of the proposed approach.