Research On SAR Image Segmentation Algorithms With Multi-scene

Synthetic aperture radar(SAR)is a device that can achieve ground imaging at different time periods and under various climatic condition.In addition,SAR also has the characteristics of long-distance imaging,strong resistance to environmental interference,and so on.Therefore,SAR has a wide range of application fields.The segmentation of SAR images is a critical component of the SAR application process,and is of paramount importance for the subsequent processing and analysis.SAR image segmentation is typically based on some characteristic information of the image to separate the image into several independent and distinct sub-regions by segmentation algorithms,in order to achieve the final good segmentation visual effect.At present,many SAR image segmentation algorithms will be studied for a specific type of scene,while SAR image segmentation algorithms in multiple scenarios will be attempted to study and analyzed in this thesis.There are three algorithms mainly examined in this thesis,which are the threshold segmentation method,the fuzzy clustering segmentation method and the Markov random field model segmentation method.Besides,these three algorithms are improved in this article,and the algorithms are introduced as follows:(1)The proposed algorithm,based on region smoothing and genetic algorithms,is improved upon the two-dimensional Otsu threshold segmentation algorithm.Firstly,the homogeneous region smoothing and edge region smoothing are carried out for the strong speckle noise in SAR images,and then the two types of smoothed images are fused to obtain the final smooth image.The threshold method proposed in this thesis is applied to segment the smoothed image,producing the final segmentation result map.Experimental results have demonstrated that the algorithm is capable of segmenting SAR images effectively.(2)The proposed algorithm,which is an improved fuzzy clustering segmentation method,is based on superpixel constrain.The FCM＿SS algorithm is put forward in this thesis,which is improved upon the FCM＿S and FCM＿S1 algorithms.In this method,superpixel pre-segmentation of SAR images is initially done,with pre-segmented images being averaged in the same superpixel block.Subsequently,the processed image is incorporated into the FCM＿SS algorithm,which performs clustering to achieve image segmentation and generate the final segmentation result map.Experiments on the segmentation of both synthetic and real SAR images demonstrate this algorithm has higher segmentation accuracy compared to other three clustering algorithms.(3)A segmentation algorithm based on nonsubsampled shearlet transform(NSST)is presented,which is based on the Markov random field(MRF)model segmentation algorithm.To begin,the input image is performed NSST to obtain high-frequency subband signals and low-frequency subband signal,and then MRF segmentation method is performed on the low-frequency subband signal to obtain the initial segmentation result map.Finally the high-frequency subbands signals are used to further fine segmentation of the initial segmentation result map to obtain the final segmentation result map.