Abstract: With the rapid development and wide application of remote sensing images, the extraction of buildings from remote sensing images can extract building information timely and accurately, which has important research significance in some applications such as rapid map updating and city management. At present, some problems such as image blurring and the improper classification of buildings exist in the gray scale image of feature map. The gray scale images need to be transformed into binary images before it can be used for the subsequent work. In order to improve classification accuracy, based on the preliminary extraction of neural network, this paper first uses the OTSU method to segment the gray scale images and uses morphological methods to process gray scale images. And it improves the Markov Random Field method and proposes a new method that can dynamically estimate the prior parameter β according to the local neighborhood characteristics of the images, introducing the original image characteristics into the Markov Random Field, further segmenting the results which is produced by the segmentation by the OTSU method and correcting the jagged boundaries of building edges in the images to improve classification accuracy. The experimental results demonstrate that the method can effectively reduce the building areas which are classified wrongly in the gray scale images extracted by neural network.

Key words: deep learning； remote sensing images； OTSU； morphology； MRF, building extraction