| With the development of imaging equipment and computer science, imagesegmentation has become a key step for computer vision systems.This paper takesnature-scene images as its subject, has done further researches on visual attention modeland Graph Cut segmentation algorithm, achieved the following results:(1) Due to the a variety of factors, the outdoor natural scenes have unstructured,diversity, randomness and complexity characteristics, resulting in image analysis andprocessing of natural scenes more difficult, while the relative stability of the visualattention model based on visual psychology mechanism，we transform the image to theHSV color space first, calculate its saliency map by hue, saturation, and other warmfeatures to split where the candidate rectangular area the object region is in, finally weakenthe impact of these complex factors in natural scenes. In order to accelerate the follow-upto the Graph Cut segmentation time-consuming, and to ensure the integrity of the object,we do further expansion to the rectangle significant area, and take the expandedrectangular area as the input image of next stage.(2) The traditional Graph Cut algorithm took image pixels as nodes to constructweighted graph, resulting that the similarity matrix is proportional to image resolution,which greatly affects the algorithm efficiency, therefore, we do improvement to thetraditional Watershed Algorithm. The method analyzes the complexity characteristics ofnatural scene images first, then smooth the original image to weaken the impact of noise,and take the distance between the edge pixels and non-edge pixels as a constraint toreconstruct the gradient map the Watershed Algorithm. The experiment result prove thatthe new method reduces the regions’ number of watershed transform effectively，at thesame time maintaining the local information.(3) In order to reduce the processing time of image segmentation, we take the pre-segmented regions (super pixels) of the improved watershed algorithm instead ofimage pixels as nodes and the similarity between regions as edges’ weights, to map thenature scene images into the weighted graph, finally take use of the max flow-min cuttheorem to complete the image segmentation. The experiment results show that thismethod not only improves the accuracy of segmentation, but also reduces the time ofimage segmentation effectively. |
