Research On SAR Image Quality Assessment And Segmentation Algorithm

Synthetic aperture radar(SAR)is not influenced by the weather and the illumination conditions.It can reconnoiter the target of interest with all-weather and all-time,and find an underground target through the mask.It could provide the ground surveying data and images with high resolution in detail even in a bad environment.Therefore,SAR has been widely used in military and civilian fields recently.With the rapid development of SAR system,the SAR image interpretation technology needs to improve urgently.The image quality assessment and the image segmentation of SAR are the essential parts of the SAR image interpretation technology.The former one is the foundation of all the applications of SAR image,including SAR image segmentation.On one hand,SAR image quality assessment technology belongs to the preprocessing part of SAR image segmentation,and is the premise of good segmentation effect;on the other hand,quality assessment can be used to evaluate the merits of the segmentation algorithm and provides the objective evaluation index for different segmentation algorithm.The research of the two could improve and optimize SAR system and establish a foundation for the automatic target recognition.The research of the two will lay the important positon in the intellectualized spatial information field for SAR system and promote SAR in application and development in many areas.In this dissertation,we research two issues:the image quality assessment and the image segmentation of SAR.The quality assessment of point target in various background clutter and the SAR jamming image quality assessment are discussed among the first part.In the segmentation section,we start with the triplet Markov model in spatial domain for SAR image segmentation,an unsupervised algorithm based on fuzzy theory with triplet Markov model is proposed.And then,from the perspective of multiresolution,two kinds of multiscale triplet Markov models based on wavelet theory are talked about for textural SAR images.Specific as follows:1.Due to the high performance and the high standards of SAR system,a more accurate evaluation result is needed for feedback to guide system adjustment.However,the traditional quality assessment experiment methods of point target usually analyze the ideal point target impulse response function.In this paper,we propose a method of point target quality assessment considering the clutter in the background.First of all,various kinds of clutter blocks are captured in the real SAR image.And then add the clutter blocks on the ideal point target image.Finally,analyze the peak side lobe ratio through impulse response function experiment method.Compared with the stochastic model,the proposed method obtains a closer result to the real measurement.As a way of experimental research,our method can provide more effective and accurate reference feedback data for SAR system.It has a certain significance.2.By studying the existing quality assessment methods of SAR jamming image,we find them lacking of perceptual.Focus on the jammed SAR image with the active noise,a perceptual quality evaluation method based on texture of SAR jamming image is proposed.On the basis of structural similarity,the new method uses the texture feature weighting the structure factor to highlight the texture feature of SAR image and obtains a measurement of structural similarity based on texture.Then,add the textural structural similarity of all the sub-bands with different directions according with the frequency characteristic between the wavelet and the contrast sensitivity function of human visual system.On this way,the effect of the medium and the low frequencies of the SAR image can be emphasized.At last,normalize the result to get a quantitatively assessment of SAR image which has a perceptibility.The experiments shows that the proposed method is closer to the subjective evaluation result than the other methods for the active noise,especially for the local jamming noise.3.On the research of SAR image segmentation,the statistical distribution models of SAR image are firstly discussed which gives a theoretical foundation to Gamma modeling.After that,we introduce a triplet Markov model in spatial domain for the nonstationary SAR image,and combine the triplet Markov model with fuzzy theory.The auxiliary field is initialized with the textural features of SAR image.And the hypothesis condition in the primary likelihood of observed field is released.These changes can improve the existing triplet Markov model segmentation method of SAR image.Use the constructed fuzzy objective function to estimate the membership and the statistical model parameters of observed field.For the parameter estimation of triplet Markov model,introduce the membership as the weighted value to adjust and update the model parameter based on the original estimation method which combine the stochastic gradient process with the iterative condition model.Thus,the purpose of optimization is achieved.The proposed algorithm can provide unsupervised SAR image segmentation,and its effect has a greater improvement than the other algorithms.4.According to the thought of multiscale wavelet domain,two triplet Markov models in multiscale are built.One of them is based on the hidden Markov tree model,and another is based on the Gaussian Markov model.Since the advantage of the interdependent relationship in the multiscale space structure of wavelet domain,the new models have the capacity for modeling the observed field more accurately.We give a new multiscale energy function for the two models at the same time,and establish multiscale potential energy of label fields.The multiscale decision fusion is realized in the process of segmentation.In the simulation experiments of synthesis and real SAR images,the proposed algorithms receive a better segmentation effect compared with the other methods.In addition,we use the maximization of the posterior marginals criterion to realize the precise segmentation for the textural SAR images in the model of triplet Markov based on Gaussian Markov,the result is more accurate.