Research On Typical Target Segmentation Algorithm Of Remote Sensing Image Based On Active Contour Model

Remote sensing image segmentation is the key technique in remote sensing image analysis and processing,because effective segmentation results can provide key information for subsequent advanced processes,such as remote sensing image analysis and recognition.Over the years,research on remote sensing image segmentation at home and abroad has gained rich achievements.Remote sensing images have statistical distribution characteristics.Therefore,level set models based on statistical distributions(e.g.lognormal distribution,Gaussian distribution,Gamma distribution,G0 distribution)are widely used in the segmentation of remote sensing images.However,the traditional active contour model based on statistical distribution can only segment the target area of remote sensing image with simple distribution.For more complex remote sensing images,these models still face various challenges.First,the solution of the energy functional should avoid falling into the local minimum,otherwise the target region will not be completely segmented,and the algorithm will be sensitive to the initial contour position.Second,these models are not suitable for the remote sensing images with complex background,strong speckle noise and blurred target boundary.Third,for remote sensing images with large image sizes,the above methods have low computational efficiency.Therefore,an effective level set model needs to be established to fit the complex remote sensing images.The specific works of this paper are as follows:(1)Theoretical studies on statistical distribution-based(e.g.lognormal distribution,Gaussian distribution,Gamma distribution,G0 distribution)level set models have been carried out.Also,a number of experiments have been done.The results show that Gaussian distribution,Lognormal distribution cannot fit the remote sensing images well,so the target area in the remote sensing image cannot be effectively segmented.The G0 distribution can relatively well fit the remote sensing images.The level set model based on the G0 distribution can segment the target area of the remote sensing images well.However,for different remote sensing images,especially remote sensing images with complex terrain distribution,the model still has the above problems.(2)This paper proposes a global minimization level set model based on G0 distribution.The G0 distribution is used to fit the intensity statistical features of the target region and the background region of the remote sensing images,then the energy functional is constructed.To avoid the local minimum problem,the energy functional is transferred into a strictly convex model that guarantees the existence of the global minimum.The theoretical basis and experiments prove that the model is not affected by the initial contour curve and it overcomes the problem that the traditional level set models based on statistical distribution are easy to fall into local minima.(3)This paper proposes a new type of sinusoidal SPF distribution.In order to enable the model to segment the remote sensing images with complex background texture,strong speckle noise and blurred target boundary,a new sinusoidal SPF distribution is introduced.The SPF function can enhance the acquisition ability to the target contour.In addition,compared with the existing linear SPF distribution,the sinusoidal SPF distribution is more advantageous in the convergence speed of the contour curve,because the rate of change of the newly constructed sinusoidal SPF distribution is relatively large.(4)This paper proposes a global minimization segmentation model based on sinusoidal SPF distribution and level set model.The newly constructed sinusoidal SPF distribution is integrated into the global minimization level set model based on G0 distribution,thus,the global minimization segmentation model based on sinusoidal SPF distribution and level set model is obtained.The theoretical basis and multiple sets of experiments prove that the model overcomes the intrinsic characteristics of the inherent multiplicative speckle noise and the heterogeneity of the target intensity in the remote sensing images.Meanwhile,the problem that the traditional level set models based on statistical distribution are easy to fall into the local minimum is solved.The evolution efficiency of the contour curve is also improved,as well as the accuracy of image segmentation.(5)In order to prove the excellent performance of the proposed method in various cases,this paper provides multiple sets of experiments using simulation data and real remote sensing images.The real remote sensing image datasets include TerraSAR-X image dataset and Sentinel-1 image dataset.