Research On Sar Image Target Configuration Recognition Method Based On NMF

Because Synthetic Aperture Radar(SAR) is able to work all day and all weather, it has been one of the important methods for earth observation and military detection. As one of the key technologies of SAR image analysis and interpretation, SAR target configuration recognition has strong commercial and military value and has become a hot spot in the world. Feature extraction is one of the critical technologies in SAR target configuration recognition. The main purpose of image feature extraction is to suppress speckle noise which would influence the recognition rate and to maximum sparse characteristic of itself. So the quality of feature extraction will directly affect the identification accuracy.At present SAR target configuration recognition methods are mostly based on intensity-based correlation matching a, two dimensional invariant moment features, or the target edge detection method, etc. These methods construct characteristic matrix by ways of extracting the coefficient of image domain or the wavelet domain. Although these methods based on global feature can obtain better recognition accuracy, their practicality is not strong for their high operation cost, slow rate and being sensitive to noise. In this thesis NMF(Nonnegative Matrix Factorization) is studied. NMF can be implemented easily and fast, and it also has definite physical meaning, which make it being one of the important research directions in the field of high-dimensional data dimension reduction analysis. In this thesis, based on the study of the existing NMF algorithm, we put forward the following three improvement methods:1. A sparse constrained non-negative matrix factorization is proposed. This method takes full use of sparse characteristic of itself and extracts sparse feature based on improved NMF. Our method is more effective to describe sparse feature of SAR image and superior to the NMF and existing sparse NMF in terms of features sparsity and characteristic chart.2. A nonnegative matrix factorization with approximate orthogonal constraints is proposed. Because of the nonnegative constraints of NMF, orthogonal constraint will bring sparse characteristic and can extract sparse characteristic efficiently. Our method ensures the nonnegative and localization of low-dimensional features, reduces the error and improves the sparse adjustment ability.3. A smooth constrained sparse nonnegative matrix factorization is proposed. This method joins the smooth constraint to the sparse nonnegative matrix factorization. Because each column of matrix has no contact and influence, the change of one image would not affect other image information. This feature is similar to Markov random process, so we join the MRF model in the process of feature extraction. This method is better than the existing sparse NMF in the sparse features, characteristic chart.