| Synthetic Aperture Radar(SAR)is one kind of active microwave imaging radar.Unlike the optical imaging system,SAR has a certain penetrating power on the ground surface or vegetation,it is capable of capturing the target information which is covered.SAR is not subject to illumination and climate,can achieve all-day,all-weather observation of the earth,and SAR also features multi-band and multi-polarization imaging.Because of these characters,SAR has a wide range of applications in the civil and military fields,in military fields,SAR automatic target recognition is a very important military reconnaissance means in the battlefield,SAR image target classification is a key technology in target recognition,it will directly affect whether the target information can be obtained accurately,so it is important to study SAR image target classification technology.For SAR image target classification problem,this thesis researched the SAR image target feature extracting method and feature coding method of SAR automatic target recognition system,achieved a SAR image target classification method based on SAR-SIFT and sparse coding.The main work of this thesis as listed below:SAR image target feature extracting densely.We adopted SAR-SIFT to extract feature from SAR image,SAR-SIFT is one modified algorithm of SIFT.SAR-SIFT improved the gradient computation method,gradient based on ratio of exponentially weighted averages takes the place of gradient based on difference,it can reduce the impact of speckle noise effectively,the extracted feature has strong robustness;For image classification problem,we extend SAR-SIFT to dense SAR-SIFT,extracting feature from the image densely;Combined with Spatial Pyramid Matching method,obtaining a fixed length feature vector,using SVM classifier to classify the SAR image target.SAR image feature coding.Sc SPM is one kind of modified method of SPM,it adopted sparse coding to take the place of vector quantization,in our research,we find that it has a shortcoming that after the sparse coding,the similar feature lost the similarity,it reduce the classification performance directly.To overcome this problem,in this thesis we modify the sparse coding method,add a non-negative local spatial constraint to the coding procedure,similar features activate similar atom sets of the codebook,so the feature codes corresponding to the similar features are also similar.The experiment result shows that by adopting non-negative local spatial constraint sparse coding method,the coding procedure is faster,and the SAR target classification precision is higher.In a further work of feature coding,inspired by the kernel method,we introduced the kernel method,kernel method maps the low dimensional feature space to the high dimensional feature space,and solves the problem that feature cannot be represent linearly in the original feature space,meanwhile in this thesis we proposed a kernel local constraint sparse coding method,maintain the similarity of the similar feature in high dimensional feature space,the SAR image target classification experiment shows the effectiveness of the method. |
