Research On Remote Sensing Image Classification Based On Texton

Remote sensing image classification is an important part in 3D scene modeling, and provides a guarantee for accurate navigation of aircraft navigation. However, for natural scene remote sensing image, due to the uncertainty of the surface features types, the morphological diversity of the same type surface feature, as well as the complexity of spatial distribution, it is still a difficult and hard problem to accurately implement surface features classification in remote sensing image.In this paper, we divide the remote sensing image classification method into two stages of image segmentation and texture recognition to carry out research on classification methods of complex nature scene from texture feature extraction, texture modeling and representation perspective.In the stage of image segmentation, for the insufficiency of image representation via single pixel in SLIC super-pixel, we use the local neighborhood information and propose a super-pixel segmentation method based on random projection(RP-SLIC), which has been reduced dimension by random projection(RP) RP-SLIC can not only reserve the advantages of compact structure and strong homogeneity of SLIC, but also better reflect the real edge contour information of the object, promote the segmentation accuracy effectively, provide a good input for the subsequent texture recognition.In the stage of texture recognition, considering the effect of illumination changes on scene imaging, we propose a novel texture feature called neighborhood difference(ND), which has some adaption for intensity change. And using ND feature, we develop two texture recognition methods based on bag of words(BoW) and Markov random field(MRF), called ND-BoW and ND-MRF, respectively.At last, we use natural scene remote sensing image to evaluate the performance of the classification method we proposed. Results show that the remote sensing image classification method in this paper can effectively improve the accuracy of image classification, with some reliability and adaptability.