An Improved Multi-scale Probabilistic Edge Detection for Urban remote sensing images

The emergence of high-resolution remote sensing images has expanded our view of urban observation and enabled us to obtain extremely rich information. However, how to detect and apply these data and make it possible to convert them into urgently needed information has become a challenge. To achieve the efficient and stable detect performance, a novel edge detection approach is proposed based on multi-scale segmentation of superpixel and the improved probabilistic evaluation of candidate contour set. In this method, the superpixel blocks have segmented firstly to reduce the number of pixels in the candidate contour. By setting different thresholds of segmentation, the multi-scale mechanism is simulated to determine the collection of pixels in candidate contour. Then the morphological method is used to extract noise pixels from the above collection. Finally, a weighted multiscale probability of boundary (mPb) index is constructed to measure and determine the candidate contour set. According to an adjustable threshold, the object is converted to a binary image in order to complete contour extraction. The effectiveness of the proposed approach is demonstrated on Berkeley benchmark data sets. To compare with several existing models, the experimental results show the advantages of multi-scale segmentation of superpixel and confirm the effectiveness and efficiency of the edge set obtained under different evaluations.