An Energy Segmentation Method of High-resolution SAR Image Based on Multiple Features

To achieve the optimal image segmentation, an energy segmentation method based on multiple features and blocks for high resolution Synthetic Aperture Radar (SAR) image is proposed in this paper. First of all, a feature vector of pixel is formed with the texture feature extracted by curvelet transform and means function, the boundary feature extracted by curvelet transform and Canny Operator, and the original spectral feature; a feature set is formed by all feature vectors of pixels in the image. The feature vector is considered as segmentation basis, and its domain is partitioned by regular tessellation. On the partitioned image domain, a label variable is assigned to a regular block; each homogeneous region is fitted by one or more regular blocks; Obviously, a label field is formed by all the label variables of regular blocks. The model of label field is built by using energy function of neighborhood relationship. The feature set is considered as a realization of a random filed of multiple features (multiple features field for short). A heterogeneous energy function is used to establish the model of multiple features field. Then the established models of the label field and multiple features field are combined to define global energy function of image segmentation, and non-constrained Gibbs probability distribution is used to describe the global energy function to build the energy segmentation model based on multiple features. Further, a RJMCMC algorithm is designed to simulate from the model to segment SAR image. To verify the feasibility and superiority of the proposed approach, real SAR images are tested.


INTRODUCTION
Image segmentation is a procedure of partitioning a digital image into some different meaningful regions with homogeneous characteristics (Hehmati et al., 2016). Synthetic Aperture Radar (SAR) image contains rich features (such as spectral, texture, boundary features), they are very important for image segmentation. However, traditional SAR image segmentation method based on spectral feature can not segment SAR image well. The reasons derive from the inherent speckle noise of SAR image, largely differences between homogeneous regions, and small differences between heterogeneous regions.
In order to improve the accuracy of image segmentation, many segmentation methods based on multiple features are proposed (Muneeswaran et al., 2006;Yang et al., 2014;Patel et al., 2011;Li et al., 2015). These methods use the advantages of different image features to segment image well.
With the development of multiresolution analysis, the image procedure methods based on multiresolution analysis are proposed (Chang and Kuo, 1993;Jung, 2007;Zheng et al., 2012). The proposed methods use multiresolution analysis to extract multiple features, and bring them into image segmentation. Wan et al. (2011)  The rest of this paper is organized as follows. In Section 2 we present the proposed algorithm. We then in Section 3 detail and discuss the results of real SAR images. Finally, Section 4 contains conclusions and perspectives for further research.

Curvelet Transform
Candes et al (2006)  [ , ] f n n ; 2) For each scale j and angle l, form the product 3) Wrapping the product around the origin and obtain collecting the discrete coefficients C D (j, l, k).

The Proposed Algorithm
Spectral, texture and boundary features are three basic features of SAR image, there are extracted and used to form a feature set in this paper, where spectral feature is select the original intensity feature, that is, z = {z d , dD}; texture and boundary features are used curvelet coefficients obtained by curvelet transform.
The operation of extracting texture feature is as follows. First, the curvelet transform is used to decompose a sub-image with c t  c t pixels centered on pixel s to obtain a series of curvelet coefficients. Then their means is computed as , , where N t is the total number of curvelet coefficients. In a ccordance with the above procedure, the t-exture feature of the SAR image is obtained, that is,  = { s ; s = 1, …,

S}.
The operation of extracting boundary feature is as follows. Firstly, the curvelet transform is used to decompose the SAR image f to obtain a series of curvelet coefficients. Then, a Canny operator is just used for extracting the boundary of curvelet coefficients on the coarse and finest scales, their non-boundary curvelet coefficients are changed to 0, and the detail coefficients are unchanged. Finally, all the curvelet coefficients are reconstructed to obtain the boundary feature of the SAR image b = {b s ; s = 1, …, S}, where  is spatial interaction parameter, NP i is the set of the neighbor blocks of P i ; l i = l i , if ( l i , l i ) = 1, otherwise, ( l i , l i ) = 0.
Then U f (f, l), represents the relationship between the characteristic field F and label field L, is defined using the sum of heterogenous potential energy function, where, f st = {f si ; l i = t}, V(f si , f st ), represents the heterogenous potential energy function, is defined by K-S distance in multiscale texture feature images (Kervrrann and Heitz, 1995) ) In order to achieve the optimal image segmentation, a RJMCMC algorithm is designed to simulate the segmentation model. In the processing of simulation, two moves are designed, involving updating label field, and splitting or merging the blocks.
1) Updating label field: the specific operation of the move is: a block P i is randomly chosen, proposing a new real label l i * in {1, …, k} (l i * ≠ l i ) and performing a switch if the proposal is accepted. The acceptance probability can be written as (Green, 1995) } ,

EXPERIMENTAL RESULTS AND DISCUSSION
To illustrate the feasibility and superiority of the proposed approach, two real SAR images with 128 128 pixels shown in  The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W10, 2020 International Conference on Geomatics in the Big Data Era (ICGBD), 15-17 November 2019, Guilin, Guangxi, China outlines on the original images shown in Figure 3. From Figure   3, it can be seen that the extracted outlines of segmentation results match the real outlines well.
In order to assess the proposed method quantitatively, some common measures, including producer's accuracy, user's accuracy, overall accuracy and kappa coefficient, are computer based on the simulated real SAR images in Figure 4. Table 1 lists them. As is shown in Table 1, it can be seen that the overall accuracies are greater than or equal to 93.7%. In addition, the kappa coefficients are up to 0.900. It can be illustrated that the proposed approach is feasible and effective.  The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W10, 2020 International Conference on Geomatics in the Big Data Era (ICGBD), 15-17 November 2019, Guilin, Guangxi, China