Fast Algorithm Based On Superpixel-level Conditional Triplet Markov Field For SAR Image Segmentation

Synthetic aperture radar(SAR) is a high-resolution imaging radar, and has the advantages of all-day and all-weather working and efficiently identifying camouflage. It has been widely used in agriculture, military and ocean areas, and has broad application prospects and great development potential. SAR image often contains a variety of information about the ground targets. It has great significance for SAR image interpretation that how to segment the all targets accurately and effectively in SAR image.SAR image segmentation plays an important role in the SAR image interpretation and is also the key and difficult point in the current research field of SAR remote sensing. As the imaging mechanism of SAR determines the introduction of a large amount of multiplicative speckle noise inevitably, the segmentation methods for the optical image can hardly achieve good results on SAR images. In recent years, the development of the theory of the random field has opened up a new path for SAR image segmentation. This paper focuses on how to obtain an effective and high efficient SAR segmentation results, and puts forward a superpixel-level conditional triplet Markov field(SL-CTMF) for the fast SAR image segmentation. The main work and contributions are as follows:1. Conditional random field(CRF) can model the posterior distribution of image directly. However, its application is limited in SAR image segmentation for the lack of effective training data and training mechanism. Triplet Markov random field(TMF) has introduced an auxiliary field U to effectively describe the nonstationarity of SAR images, such that it can suppress the influence of multiplicative speckle noise to SAR image segmentation and good segmentation results have achieved. However, the modeling of TMF is complex and TMF cannot make full use of the correlation of the observed data.2. It is the idea that CRF directly models the posterior of the image to solve the shortcomings of TMF model. Hence, the pixel-level conditional triplet Markov field(PL-CTMF) model had been presented. This model fully combines the advantages of CRF and TMF: modeling the posterior probability of the X field directly andintroducing the auxiliary field U to describe nonstationarity of SAR image. PL-CTMF simplifies the method to model the SAR image and improves the SAR image segmentation.3. However, in PL-CTMF no matter how similar the pixel’s feature to its neighborhoods’ is, the probabilities of all pixels in the image are still needed to estimate. The low efficiency and high redundancy are inevitable. To achieve an effective and high efficient SAR image segmentation, this paper proposes the SL-CTMF model. Firstly, for SAR image, we improve the Turbo Pixels algorithm to make it has the ability to obtain a superpixel-level SAR image in accurate edge location. Then based on the superpixel-level SAR image, an auxiliary field U is reconstructed to describe the nonstationarity of the SAR image. The unary and pairwise potential of the SL-CTMF are also derived with the superpixel-level feature and texture information. As the SL-CTMF is to label these superpixels and the feature of each superpixel is the integrated feature of all the belonging pixels, the efficiency and regional consistency of the segmentation result can be effectively improved. Finally, the SL-CTMF is applied to fast unsupervised real SAR image segmentation with the maximum posterior marginal(MPM) method. Since the segmentation result of SL-CTMF is similar to or a bit better than PL-CTMF, the running time of the SL-CTMF is much shorter than PL-CTMF, as it is 1/4 to 1/6 of PL-CTMF’s.