Abstract: For the classic wavelet textures, the image texture features cannot be accurately expressed. The image segmentation results lack the consideration of the spatial correlation and distribution relationship of the pixels. In this paper, we proposes a combination of dual-tree complex wavelet (DT-CWT) texture and Markov random field (MRF) model for high resolution remote sensing image segmentation method. Firstly, the image texture feature is extracted by dual-tree complex wavelet transform, and the mixed feature vector of the expression image is formed by combining the texture feature and the spectral feature. Then, Gaussian normalization of the mixed feature vectors in the feature space is performed. The K-means clustering method is used to perform the feature vectors in the feature space. The initial segmentation map is obtained by clustering. Finally, using the Markov random field model to introduce context information to represent the initial segmentation results, the initial segmentation results are optimized based on the Bayesian maximum a posteriori probability criterion, and the final segmentation results are obtained. This paper improves the accuracy of feature expression by using dual-tree complex wavelet textures. At the same time, the Markov random field model is used to weaken the "pepper and salt noise" in the homogenous region of the segmentation results, further improving the segmentation accuracy of the high resolution remote sensing image.

Key words: dual-tree complex wavelet transform(DT-CWT), texture feature, Gaussian normalization, Markov random field, high resolution remote sensing image