Research On Occluded Target Recognition Of Synthetic Aperture Radar Based On Machine Learning

Synthetic aperture radar(SAR)plays an important role in both the military field and the civilian field.As an important aspect of the application of the SAR,target recognition always is the research hotspot.The target recognition methods based on the machine learning have achieved great performance.The current target recognition methods take the complete target as the processing object.In practical application,the target can be incomplete due to the occlusion.The performance of target recognition based on the traditional machine learning for the occluded target may be seriously deteriorated.Therefore,it is necessary to study the target recognition method for the SAR occluded target.This paper focuses on the recognition method of the SAR occluded target based on sparse representation(SR)and convolutional neural network(CNN).The main work of this paper is shown below.1.The adaptive weighing model based on sub-image sparse model for the SAR occluded target recognition is proposed.The key of the occluded target recognition based on sparse representation model is to eliminate the influence of the occlusion information on sparse reconstruction error.The proposed method splits the test sample and the dictionary into the sub-test samples and the sub-dictionaries.The sub-dictionary is used to sparsely represent the sub-test sample and calculate sparse reconstruction error of the sub-test sample.The sub-test samples with large sparse reconstruction error are treated as the sub-test samples containing occlusion information.The sub-test samples containing the occlusion information and the corresponding sub-dictionaries are weighted to obtain the weighted test sample and the weighted dictionary.The weighted dictionary is used to sparsely represent the weighted test sample and recognize the target.The performance of the proposed method for the unoccluded target is similar to the traditional sparse representation method.The performance of the proposed method for the occluded target is better than the traditional sparse representation and the support vector machine.2.The fusion strategy based on the weighed image for the SAR occluded target recognition is proposed.The key parameters of the weighted image model include the sub-image size,the weighted sub-image ratio,and the weight.The values of these parameters directly affect the recognition performance of the method.It is very difficult to set the optimal parameter values of the weighted image model.The proposed method uses the fusion strategy to overcome this problem.The proposed method sets a series of parameter values to obtain multiple weighted test samples and multiple weighted dictionaries.Then,the pixel-level fusion or the decision-level fusion is selected to process these weighted test samples and the weighted dictionaries.Finally,the fusion result is used to recognize the target.The proposed method avoids setting optimal values of key parameters in the weighted image model.The recognition performance for the occluded target has been improved.Experimental results show that the proposed method has better recognition performance than traditional target recognition methods.3.The fusion model of sparse model based on randomly erased image for the SAR occluded target recognition is proposed.It is often difficult to accurately locate the occlusion information in the sparse representation model.The proposed method uses the random erasure to circumvent this problem.The proposed method randomly selects an area in the test sample and the dictionary and sets the values of these pixels to zero.The zeroed area of the test sample is the same as the zeroed area of the dictionary.The proposed method uses the decision-level fusion strategy to overcome the under-erasure of the occlusion information and the false erasure of the target information in random erasure.First,the test sample and the dictionary are randomly erased multiple times.Then,the erased dictionary is used to sparsely represent the corresponding erased test sample and calculate the sparse reconstruction error.Finally,the proposed method fuses these sparse reconstruction errors.The target is recognized based on the fusion result.The proposed method can eliminate the occlusion information and retain the target information.Experimental results show that the proposed method has better recognition performance than traditional target recognition methods.4.The modular construction method of the CNN and the recognition method of the SAR occluded target based on data augmentation are proposed.The topology of the CNN directly determines the performance of the network.The proposed method uses the modular idea to simplify the design of the CNN.First,the proposed method builds the basic module.The basic module includes the convolution layers and the pooling layers,which is responsible for the function of feature learning and dimension reduction.Then,the CNN is obtained by stacking the basic module.The CNN obtained by the proposed method has similar performance to the CNN designed by the traditional method.In order to improve the performance of the CNN in recognizing the occluded target,the proposed method uses the data augmentation to ensure that the CNN learns the characteristics of the occluded target.The data augmentation includes static data augmentation and dynamic data augmentation.The training samples of the occluded target in static data augmentation are fixed.The training samples of the occluded target in dynamic data augmentation are changing.Experimental results show that the data augmentation can significantly improve the performance of the CNN to recognize the occluded target.5.The multi-level sparse representation model based on reconstruction error level for the SAR target recognition is proposed.The sparse representation model uses the sparse reconstruction error as the classification index.Ideally,only one sparse reconstruction error on the sub-dictionary is small,and the sparse reconstruction errors on the other sub-dictionaries are large.The proposed method uses the sparse reconstruction error to represent the representation ability of the dictionary and determine the processing method of the test sample.The proposed method sets two thresholds.The first threshold represents the absolute representation ability of the dictionary and is used to determine the processing flow of the test sample.The second threshold represents the relative difference of the sparse representation ability of these sub-dictionaries,and is used to decide whether to reconstruct the dictionary.Finally,the proposed method uses the recognized samples to improve the dictionary.This method further explores the property of the sparse reconstruction error and achieves better recognition performance.6.The multi-view tensor sparse representation model for the SAR target recognition is proposed.The tensor sparse representation model can retain the local structural information of the SAR image.The SAR can obtain multiple images by observing the target multiple times.These images have internal correlation.The proposed method proposes the multi-view tensor sparse representation model to comprehensively utilize the local structural information and the internal correlations of these SAR images.The proposed method proposes the joint tensor orthogonal matching pursuit algorithm to calculate the sparse coefficient tensor of multiple views.First,the proposed method uses the classic tensor dictionary learning method to obtain the dictionaries of all classes.Then,the proposed method uses the joint tensor orthogonal matching pursuit algorithm to calculate the sparse coefficient tensors of multiple views.Finally,the target is recognized based on the total sparse reconstruction error of all views.The performance of the proposed method is better than that of the joint sparse representation model and the sparse representation fusion model.