Automatic Segmentation of Background Defocused Nature Image

Fully automatic segmentation of natural image is still a tough task. In this paper, we investigate a special class of natural images, whose foreground objects appear sharp while background are blurred due to out of depth of field, we call that Background Defocused Images. This kind of pictures is frequently seen on newspapers, advertisements and feature shoots. An algorithm is proposed to automatically segment the foreground from the background in pixel level. First, a novel local blur measuring method is proposed, which is based on a precise edge width calculation algorithm. Then, Graph cut(s-t cut) algorithm is used to find a further segmentation both for undetermined area and for smoothing. In experiments with large foreground blur, object can be captured in a good precision. This technique can be used in image editing, or as a part of automatic annotation, segmentation or recognition systems.