Research On SAR Image Denoising Algorithm Based On Dictionary Learning

Synthetic aperture radar is a kind| of remote sensing imaging radar,which can collect the realtime data on target area.The advantage of highly efficiency of imaging in poor conditions makes the SAR is widely applied in various fields,such as disaster monitoring,terrain mapping,resource exploration and military reconnaissance.However,due to coherence from distributed targets,SAR images suffer from the speckle noise,which mixes with small targets and carries itself about the image information,seriously distorting the visual interpretation and subsequent processing of images.However,denoised images have a tendency that some regions in images are blurred by filtering methods.It is significant to develop despeckling methods that can keep an acceptable balance between despeckling and edge preserving.However,with the constant improvement of linear sparse representation theory,dictionary learning has been widely practiced in image processing,but there are still some problems need to be solved in SAR images application.The content of this dissertation is focus on the despeckling method that based on the theory of dictionary learning.The major works are as follows:(1)A nonlocal self-similar priority dictionary learning algorithm for SAR image despecklingIn order to compensate for drawbacks of the traditional dictionary learning method,including denoising without similarity of images and low-efficient dictionary updating order,a nonlocal selfsimilar priority dictionary learning algorithm is proposed in this dissertation.First,the training data is collected by image patches,which are depended on the characteristic of SAR images.Second,compared with the sequential update order of traditional algorithms,the update order of priority dictionary learning calculates the number of nonzero items in every rows of the coefficient matrix in descending order.The first updating begins form the atom which has the highest value in non-zero items.Final,the next index of atom is the highest value in outputs calculated by formula which depends on the number of overlaps between non-updated rows and updated rows of coefficient matrix at the position of nonzero item.From the results of experiments,the proposed algorithm can achieve outstanding despeckling performance with the final dictionary,which has the more accurately reconstruction ability.(2)A superpixel-based clustering multi-dictionary learning method for SAR image despecklingTo solve the problem that a single dictionary is unsuitable for SAR image despeckling with the complex information,a multi-dictionary SAR image denoising algorithm based on superpixel is proposed in this dissertation.First,the similar pixel in neighbor regions is aggregated in a superpixel.Unlike traditional rectangle,a superpixel is uniform in size and shape and thus,has better flexibility to gather similar pixels.A number of superpixels contains all the pixels of the image.Second,a clustering algorithm is used to classify the superpixels into several classifications,obtaining different scattering regions.The superpixels with the same characteristics are classified into one classification.Final,initialized dictionaries are used to denoise classified areas respectively.Experimental results exhibit that the proposed algorithm can improve despeckling performance for different scattering areas of SAR images.