Superpixel Segmentation Method For Remote Sensing Image

With the rapid development of remote sensing data acquisition technology, the resolution of remote sensing image is being improved, which results in the large-scale data and increasing proportion of redundant information that restricts the data processing and analysis. The scholars of computer vision raised the concept of superpixels that means grouping the contiguous pixels having similar color and textures. Compared with pixels sampled regularly, both superpixels have visual perception but also can reduce data redundancy and data scale. However, the recent superpixel segmentation algorithms mainly aim at general pictures, not for remote sensing image, the superpixels generated by segmentation do not have the regular arrangement and shape. As for high-resolution remote sensing image, most objects have the regular geometric shapes arranging regularly. To this end, two superpixel segmentation algorithms are proposed to produce the superpixels having the regular arrangement and shape. The first algorithm segments the image into a grid structure initially, then using property of superpixels’region and boundary to construct objective function optimized by dynamic programming method to find optimum solution to adjust the initial segmentation. The algorithm will iterate until contain a set of superpixels that attached to the edge of image. The second algorithm divides the image into a grid structure initially, too. Then using edge information of image to drive the grid points in a local area to move iteratively to make the quadrilateral grid attached to the edge of image. The second algorithm can not only ensure superpixels arranged regularly, but also ensure the superpixels been constructed by four line segments. This vector form cannot only compress the data greatly, but also effectively fit the polygon target that has important applications in remote sensing image object recognition.