Based On Compression Perception Of Sar Image Target Recognition Technology Research

Synthetic Aperture Radar (SAR) could obtain two-dimensional high-resolutionSAR images of surface features terrain in the complex environment, which provides amore reliable means for earth observation and military reconnaissance. Faced with therapid expansion of SAR data sources, how to extract useful information fromlarge-scale data quickly and accurately posed a challenge to SAR image interpretationsystem. As a key technology for SAR image interpretation system, Automatic TargetRecognition (ATR) of SAR image plays an important role in information extraction,and thus became a hotspot to domestic and foreign scholars. Compressed Sensing (CS)is a new signal processing theory developed in recent years, which shows a larger valuein the field of target recognition. Combined with the features of SAR image, this thesisfocuse on the technology of SAR image target recognition based on compressed sensing,the main contents are as follows:1. Aiming at remove interfering background and enhance classifiable information,the combination of preprocessing is investigated for SAR image. Firstly, an algorithmof removing interferential clique based on Markov Random Field (MRF) is proposedfor SAR image segmentation. Secondly, for removing isolated pseudo-target area aftersegmentation, geometric clustering is applied to SAR image. Finally, imageenhancement, centroid registration and gray-scale normalization are used to target areain SAR image. Experiment results shows that the preprocessing could obtain completeoutline of the target, remove isolated pseudo-target area, and preserve the details of theedge and structure of the objectives, ultimately enhance the classifiable imformation inSAR image.2. In consideration of the high performance when applied compressed sensing topattern recognition, a new method of SAR image target recognition based oncompressed sensing is researched. In order to reconstruct the non-negative constraintssparse representation between samples, the Non-negative Gradient Projection for SparseReconstruction (NGPSR) algorithm is derived on the basis of Gradient Projection forSparse Reconstruction (GPSR) algorithm, and then based on non-negative sparse coefficient, a classification algorithm is designed for SAR image target recognition.Finally, simulation results based on MSTAR data verified that, with respect to threetypes of targets contains variants, this method could achieve a97.29%averagerecognition rate, hence it is an effective SAR image target recognition method.3. According to the rejection criteria based on the measurement of confidence, anew method which combines Nearest Neighbor (NN) and compressed sensing isproposed for SAR image target recognition. After the classification by nearest neighbor,for the samples with higher confidence, its category is directly determined by nearestneighbor classifier, for the other samples with lower confidence, first judge them to berejection samples, and then its category is determined by compressed sensing method.As utilized K Nearest Neighbor (KNN) classifier for select part of the training samplesto compose sparse dictionary in compressed sensing, the dimension of the dictionary isreduced greatly, and therefore the reconstruction speed of sparse representation isaccelerated. Experiments based on full and part training samples from MSTAR databaseboth comfirm that, in the case of slightly lower the recognition rate, this method couldobviously improve the speed of target recognition.