Optimization Of SAR Image Feature Selection Method Based On Genetic Algorithm

In order to match the interpreting capability of the radar image with the high-speed imaging capability of the image,the automatic recognition system(ATR)has risen rapidly and has been highly valued by various countries,among which,the three-stage flow proposed by Lincoln lab,including testing,discrimination and identification,is the most classic.Based on the classical three-stage process,this thesis has studied from four aspects:First,fully studied and summarized the models of clutter in Synthetic Aperture Radar(SAR)images,and selected the most representative models,G0 distribution and Rayleigh distribution,to make a comparison,which shows that the G0 distribution is more reasonable for clutter modeling of SAR images,then 789 clutter slices are successfully acquired using Constant False Alarm Rate(CFAR)algorithm;Second,summarizing various algorithms for SAR image target discrimination feature extraction,finally,based on the proposed three criteria of feature practicability,effectiveness,stability and high efficiency,the features based on contrast and dispersion are selected as the object of this study,based on the analysis of the characteristics of fractal dimension,peak-to-energy ratio,and four spatial boundary features,a method of self-adaptively selecting slice description points based on slice contrast is proposed,by comparing the noble methods,we can see that the new method improved the discrimination rate of quality characteristics and peak signal-to-noise ratio by 19.65%and 10.94%respectively.Third,studied and summarized the existing classifiers,indicated that Bayesian discriminator and Support Vector Machines(SVM)are most likely to be practical discriminators,through our experiments,it has been found that both discriminators have feature redundancy problem,thus,feature selection has to be done,because these two discriminators have almost the same discrimination effect,but the efficiency of Bayesian is much higher than that of the SVM,for practicality,we use genetic algorithm combined with Bayesian discriminator to find the optimal combination of features,which eliminates feature redundancy,at the same time,Bayesian discriminator and fitness function optimization methods are proposed,compared with the noble method,our method eliminates missed detection then greatly reduces the risk of usage,moreover,the average identification time was reduced from 9.77 seconds to 7.23 seconds.Fourth,because the distribution of discrimination features show some differences for different target models,the authors used the Bayesian principle,combined with genetic algorithms and Bayesian classifier,using nine features based on contrast and dispersion,proposed a fast target model identification method.The experimental results show that,the recognition rate has reached 86.73%on nine targets with large external differences in structure,compared with other eight classical recognition methods with an average recognition rate of 92.7%,there is no need for complex preprocessing,3 minutes fast recognition time on 1213 targets and stable recognition ability,make it possible to become an auxiliary model recognition tool.