Research On Aerial Target Recognition Based On Two-dimensional Entropy Image Processing And SVM

Target recognition technique based on image has been widely applied in many fields, now, it has penetrated into such fields as medical, industrial production, aerospace, transportation and so on. The support vector machine is a kind of novel machine learning methods to solve the nonlinear and high dimension recognition problem with small sample set. Now, it`s a standard tool in the fields of machine learning and data mining. This paper researched on aerial target recognition based on two-dimensional entropy image processing and SVM.In the paper, SVM, which is based on statistical learning theory, is applied to target classification according to aerial targets are prone to rotation, scale changes, occlusion and low contrast features. The introduction of combination features improved the classification accuracy of the algorithm.The quality of image segmentation seriously affects the subsequent feature extraction. Two-dimensional entropy thresholding segmentation,which is based on two-dimensional histogram and compartmentalization, has the problems that useful information is lost and it can not effectively eliminate the noise interference. Proposed a method of the building of two-dimensional histogram based on the gray-gradient co-occurrence matrix model. Entropy thresholding function was construction by two-dimensional Renyi\Arimoto Entropy and the new two-dimensional histogram, then gained two new image thresholding segmentation methods. It is proved in experiment that new methods have strong noise immunity and can describe the boundaries of objective well. This meet the subsequent feature extraction. New appproach, which had a strong adaptability, could get more complete edge and homogeneous inner of the target.After researching on the support vector machine model and parameter selection, normalized combination features matrix was calculated and a training model of the target was completed with the LIBSVM, then, aerial target classification and recognition system was build. The results of simulation and experiment show that proposed classification and recognition algorithm has strong robustness and accuracy.