Application Of Markov Random Field With Multi-Type Feature Coordination In Remote Sensing Image Segmentation

In recent years,with the successful launch of high spatial resolution remote sensing satellites at home and abroad,the resolution of Earth observation data is getting higher and higher.However,the ability of remote sensing image data processing in China urgently need to be strengthened for the development of high resolution remote sensing systems in the near stage.Therefore,the analysis and processing of high-resolution remote sensing data has become an important scientific issue in the field of remote sensing.Segmentation is the most important and challenging problem in this application field.For the segmentation of massive remote sensing image data,it is concerned by the method based on statistical probability.Among them,Markov Random Field(MRF)method is more widely applicable.However,there are some existing improvements in the existing methods based on it:(1)In the pixel-level feature extraction method,the neighborhood range of pixels is relatively small,so the segmentation result loses important detail features and produces.A large number of fragments.Therefore,misclassification often occurs in the segmentation process.(2)The segmentation accuracy is limited,which depends on the initial segmentation inaccuracy of the MRF model and the irregularity of the context relationship in the object-level feature extraction method.To address the above problem,the following research work has been done in this paper.The Alternating Iterative Markov method for remote sensing image segmentation(AIMRF)is proposed,which enables the model to simultaneously consider the spatial context information for the rule and the macroscopic texture description ability to promote and optimize each other.Firstly,two probability maps are established to extract pixel-based spatial information and object-based spatial information respectively.Secondly,corresponding sub-models are established on the two probability maps respectively.Finally,an interactive iterative pattern is proposed to fuse information.Continuous optimization between the pixels level feature and the object level feature results in the best segmentation result.The Pixel-level depth Markov method for remote sensing image segmentation(PDMRF)is studied,which makes the model not only have the superior structure of the convolutional neural network but also do not require a large number of training sets and can achieve better results in the region consistency and the original information description ability of image.Firstly,the convolution structure is used to extract different aspects of the original image data.Secondly,the pooling operation is performed in each dimension to realize information down-sampling of different dimensions.Then,the principal component analysis is used to extract the feature information of the obtained information.Finally,the classical pixel-level method is used to achieve the optimal segmentation of the original image.In order to verify the validity of the two models proposed in this paper,we compare with other existing model-based methods on different datasets.The evaluation indicators show that the proposed method demonstrates a superior performance in image segmentation.