Research On SAR Image Denoising Based On Low Rank Reconstruction And Component Analysis Theory

Compared with optical remote sensing,synthetic aperture radar(SAR)remote sensing can collect data without any restriction of time,region,and weather.The imaging is not affected by weather,region and climate.At the same time,the resolution of SAR image has nothing to do with the target distance.It plays a very important role in national defense and life.Because of the coherent imaging systems,speckle noise inevitably occurs in SAR images.Speckle noise is a kind of signal-dependent granular noise,which is inherent in coherent imaging system.The presence of speckle reduces image quality and affects subsequent segmentation and interpretation.Therefore,it is particularly important to research the denoising algorithms for SAR image characteristics.Low rank reconstruction and component analysis theory have important theoretical value and application potential,and have attracted much attention from scholars at home and abroad in recent years.However,there are still many problems to be solved in the application of low rank reconstruction and component analysis theory to SAR image denoising.On the basis of previous studies,SAR image denoising methods based on low rank reconstruction and component analysis are studied in this dissertation.The main innovative research results of this dissertation are as follows:(1)A blind denoising method for SAR images based on texture strength and weighted nuclear norm minimization is proposed.Firstly,the structure features of the SAR image are analyzed and characterized by the trace of the gradient covariance matrix.Thus the texture strength reference measurement representing the texture structure of the image is obtained.The low-rank patches are selected by texture strength,and the noise level of the image is estimated by the selected low-rank patches,which makes the proposed method in this section a blind denoising method.By using the estimated noise variance and the singular value of the observed noise image patches,the singular value of the unknown original noise-free image can be estimated.Considering that the larger singular values represent the principal component of the original data,they should be retained as far as possible.While the smaller singular values represent the less important component of the original data,they should be shrunk as much as possible.Therefore,different weights are given to different singular values,and the weights should be arranged in a non-descending order.Usually,the weighted nuclear norm minimization problem is a non-convex optimization problem.But when the weight coefficients are arranged in a non-descending order,the problem has an optimal analytical solution.Therefore,the weighted nuclear norm minimization can be used for image denoising and obtain the optimal solution.The experimental results show that the blind denoising method based on weighted nuclear norm minimization remains competitive in both subjective and objective aspects in processing SAR images.More importantly,the noise level of actual SAR images is usually unknown and we propose a blind denoising framework for SAR images.On the one hand,the noise variance can be estimated by using low rank patches.On the other hand,the noise variance can be adjusted with the selected denoising algorithm to guarantee the denoising algorithm achieves the best denoising result.The blind denoising method of SAR images based on texture strength and weighted nuclear norm minimization has higher robustness and good application prospects in engineering practice.(2)A boosting synthetic aperture radar image despeckling method based on non-local weighted group low-rank representation is proposed.Based on the structural redundancy in SAR images,the low-rank restoration method is used to recover the original data.For each pixel,the rank-ordered absolute difference is used to calculate its contamination probability.And the contamination probability is added to the low-rank representation model as a constraint condition to restrict the regularity of the restored image.Weighted averaging is used to aggregate each patch to further suppress noise and the boosting algorithm is used to reduce the difference between local filtering and global restoration.Then the denoised image is obtained.The experimental results show that the algorithm obtains a higher objective index in the simulation speckle noise model,especially in the aspect of background contrast,which shows that the radiation fidelity of SAR image is better protected.The local enlarged images of simulated and real SAR images show that,compared with similar methods,the method based on weighted low rank restoration has better texture protection ability while smoothing noise effectively.(3)A new SAR image denoising method based on the grouping-based principal component analysis and guided filter is proposed.We model a pixel and its nearest neighbors as a vector variable.Vectors similar to the vector to be processed are selected in the local window to form a training set.Block similarity measure is used as similarity criterion to find vectors with similar structure to the vector to be processed.The training set is analyzed by principal component analysis,and then the transformation coefficients are shrunk.By using local grouping,the local information of variables can be calculated accurately,so that the edge structure of the image can be better protected after shrinkage coefficient in PCA domain.Due to the existence of strong speckle,local vector grouping errors and PCA transformation matrix errors maybe occur,there is still noise residue after the coarse filtering.Guided filtering not only has good edge smoothing characteristics,but also can be calculated efficiently.Therefore,guided filtering is used to further denoise.The experimental results show that this method is superior to the non-local methods in protecting boundary and avoiding artifacts.(4)A SAR image denoising method combining wavelet-contourlet with iterative cycle spinning is proposed.Firstly,the shortcomings of the existing multi-scale geometric transformation are analyzed,such as lack of shift invariance and high redundancy.Considering that contourlet transformation does not have shift invariance,pseudo-Gibbs phenomenon will occur when the threshold is set for the transformed coefficients.Iterative cycle spinning is used instead of multiple shifts and averages,which can reduce the pseudo-Gibbs phenomenon better.The actual SAR image denoising experiments show that the wavelet-contourlet method combined with iterative cycle spinning method not only improves the objective performance parameters,but also has better visual effect than similar methods and reduces artifact information.