Study And Development Of Radar Target Classification Algorithms

Synthetic aperture radar(SAR)is an airborne and space-borne remote sensing system for imaging distant targets on a terrain,which can operate proficiently in all-weather day-and-night conditions and generate images of extremely high resolution.SAR images take advantage of longer-wavelength signals to provide complementary information to that of optical sensors operating in the visible and infrared regions of the electromagnetic(EM)spectrum.Moreover,unlike photographic images that contain only the amplitude information of the target?s reflectivity of light,SAR provides both the amplitude and the phase information of the scattered EM field from the scene.Therefore,using SAR for navigation has become an indispensable part of modern life and the classification of these SAR images have become one of the greatest challenges.However,the understanding of SAR images is hard to carry on manual interpretation compared to optical images which describe a good appearance of an object.In addition,SAR images are prone to be contaminated by multiplicative noise,which makes it very difficult to perform target classification in SAR images.This work is focused on the efficient classification methods of the SAR imagery.In the classification system,there are two different mainstreams;hand-designed features based approaches and deep convolutional neural networks(DCNNs)based approaches.The first part of this thesis introduces two hand-designed based feature extraction methods for SAR image target classification which are presented in chapter 3 and 4,respectively.The first classification method is based on a visual saliency model.First,a SAR oriented graph based visual saliency(GBVS)model is introduced.Second,relying on the competence of our model in highlighting the most significant locations where the target image is informative,Gabor and HOG features are extracted from the processed SAR images.Third,in order to have more discriminative features,the discrimination correlation analysis(DCA)algorithm is used for feature fusion and combination.Finally,a two-level directed acyclic graph support vector metric learning(DAG-SVML)is developed that seamlessly takes advantage of a two-level DAG by eliminating weak classifiers and the Mahalanobis distance-based radial basis function(MDRBF)kernel which emphasizes relevant features and reduces the influence of non-informative features.Experiments on Moving and Stationary Target Acquisition and Recognition(MSTAR)public release database are conducted,and the classification accuracy and time complexity results demonstrate that the proposed methods outperforms the state-of-the-art of the hand-designed methods.The second classification method introduces bag-of-visual-words(Bo VW)based feature representation.In this method,Gabor and HOG based features are adopted to extract features from the training SAR images.Then,a discriminative codebook is generated using K-means clustering algorithm.After feature encoding by computing the closest Euclidian distance,the targets are represented by new r obust bag of features.Finally,the linear support vector machine(SVM)is used as a baseline classifier.The proposed Bo VW based approach achieves better time complexity and less classification accuracy compared to the case of SAR oriented GBVS up to 22.9 ms and 0.32 %,respectively.The second part of this thesis introduces two DCNNs based approaches in chapter 5 and 6,respectively.The idea behind these approaches is to discover multiple levels of representation so that higher level features can represent the semantics of the data,which in turn can provide greater robustness to intra-class variability.The first method fine-tunes the pre-trained VGG-S Net on MSTAR database.Then,the fine-tuned network is used as a fixed feature extractor to extract deep features from the processed SAR images.After that,the extracted deep features are fused by using a traditional concatenation and a DCA algorithm.Finally,K-Nearest Neighbors algorithm(K-NN)based on Log Det Divergence-Based Metric Learning Triplet Constraints(LDMLT)is adopted as a baseline classifier.Experimental results on the MSTAR database demonstrate that the proposed method attains good results compared to the state-of-the-art methods.The second classification approach is based on very deep learning and multi-canonical correlation analysis(MCCA).This method proposes a robust feature extraction method for SAR image target classification by adaptively fusing effective features from different CNN layers.First,a very deep CNN is fine-tuned on MSTAR database by using small filters throughout the whole net to reduce the speckle noise.Besides,using small-size convolution filters decreases the number of parameters in each layer and therefore reduces computation cost as the CNN goes deeper.Moreover,Re LU layers are integrated to have more discriminative decision functions and lower computation cost.The resulting CNN model is capable of extracting very deep features from the target images without performing any noise filtering or pre-processing techniques.Second,our approach proposes to use th e MCCA to adaptively learn CNN features from different layers such that the resulting representations are highly linearly correlated and therefore can achieve better classification accuracy even if a simple linear support vector machine(SVM)is used.Experimental results on the MSTAR database demonstrate that the proposed very deep learning algorithm outperforms the state-of-the-art methods.