Sparse Feature Learning Based SAR Images Segmentation And Semi-supervised Classification

Synthetic aperture radar(SAR)can penetrate clouds and vegetation,be capable of all-weather and all-time operation,and obtain image with high resolution,therefore it has been widely applied in military and civilian fields.With the development of SAR technology,large data(such as SAR images)can be easily obtained,so SAR images understanding and perception are very important.SAR image segmentation and classification are the essential parts for subsequent SAR images understanding and perception,which has been studied by many researchers in recent years.Recently,as an effective technique to analyze the sparsity of large data,sparse representation has attracted a growing interest,and has been successfully used in many applications.Considering inherent characteristics of SAR images,this paper utilizes the advantages of sparse representation to research SAR images segmentation and classification problems.The main content of this paper is summarized as follows:1.Considering that SAR images contain speckle noise,they are firstly over-segmented into many superpixels before being segmented and classified.As a basic unit,the superpixel not only effectively reduces the speckle noise of SAR image,but also significantly improves algorithmic efficiency.To further decrease the number of superpixels and well connect the superpixels and image pixels,the coarse merging and boundary processing methods of superpixels are presented.Then,by extracting multiple features of superpixels,different objects in SAR images are fully and accurately described.2.A multi-kernel joint sparse graph(MKJS-graph)is proposed to segment and classify SAR images.At first,a new multi-kernel sparse representation model is used to express sparsely multiple features of the superpixels in high dimensional kernel space,which can reflect the global similarity of superpixels.Moreover,the local neighborhood spatial correlation of superpixels is combined with the global similarity of that to improve the segmentation performance by formulating the adjacent matrix for MKJS-graph.Integration of the global and local structures of the superpixels provides the MKJS-graph with favorable category distinguishing ability on segmenting and classifying SAR images polluted by speckle noise.The simulated and real SAR images are tested through a series of experiments,and the results indicate that the proposed method is more competitive than other state-of-the-art algorithms in SAR image segmentation and classification.3.A new random subspace based ensemble sparse representation(RS_ESR)algorithm is presented and applied in SAR image segmentation and semi-supervised classification.For high-dimensional data,random subspace method can not only reduce dimensionality of data but also make full use of effective information in data.Then a joint sparse representation model is used to get the sparse representation of sample set in each random subspace.These sparse representations in multiple random subspaces are integrated as an ensemble sparse representation by a simple way.Moreover,the obtained RS_ESR is applied in machine learning tasks,which includes classical spectral clustering and semi-supervised classification.Experimental results on different real-world data sets show the superiority of RS_ESR over traditional methods.4.A sparse learning based fuzzy c-means(SL_FCM)method is proposed.We introduce sparse representation into the fuzzy clustering to improve the clustering performance and reduce the space and computational complexity.Firstly,most of energy of discriminant feature obtained by solving a sparse representation model is reserved and the remainder is discarded.By this way,some redundant information(i.e.the correlation among samples of different classes)in the discriminant feature can be removed,which can improve the clustering quality.Furthermore,the position information of valid values in discriminant feature is also used to re-define the distance between sample and clustering center in SL_FCM.It can enhance the similarity of the samples from the same class and the difference of the samples of different classes,which contributes to the clustering.In addition,as the dimension of stored discriminant feature of each sample is different,we use set operations to formulate the distance and cluster center in SL_FCM.Experimental results manifest that SL_FCM has favorable clustering performance in low memory space and computational complexity.5.A novel fuzzy double c-means based on sparse self-representation(FDCM_SSR)algorithm is presented.The major characteristic of FDCM_SSR is that it can simultaneously address two datasets with different dimensions,and has two kinds of corresponding cluster centers.The first one is the basic feature set that represents the basic physical property of each sample itself.The second one is learned from the basic feature set by solving a spare self-representation model,referred to as discriminant feature set,which reflects the global structure of the sample set.The spare self-representation model employs dataset itself as dictionary of sparse representation.It has good category distinguishing ability,noise robustness,and data-adaptiveness,which enhance the clustering and generalization performance of FDCM_SSR.Experimental results on different datasets and images show that the performance of FDCM_SSR is better than other state-of-the-art fuzzy clustering algorithms.