Sidescan sonar image is a key step for marine object recognition and classification, and still remains a challenging task as the images formed from acoustic signal are severely interfered with noise. Moreover, its real time processing is indispensable for subsea robots. In this paper, we proposed a fast and precise method to implement real-time object segmentation. The method integrates preprocessing using a graph-based bothway spanning forest(BSF) algorithm and fine segmentation using level sets method with region-scalable fitting (RSF) model. The formal part is developed from the minimum/maximum spanning tree algorithm, which deals with sonar images efficiently and provides a coarse segmentation result as the initial input of the latter part, and then RSF model refines that result and obtains a final object segmentation. Our method achieves a faster speed by 2 s per image (14000*80) compared to the related work and a perfect segmentation accuracy, which outperforms the state-of-art method.