Study On Infrared And Hyperspectral Image Processing Based On Data Classification

Classification is one of the typical processing methods of remote sensing images and is widely used in remote sensing applications.For example,the classification of small target and background pixels in single-frame infrared image is one of the key technologies for small target extraction in applications such as infrared guidance and space-based early warning systems,whereas hyperspectral imagery(HSI)classification is widely used in remote sensing quantitative observation applications,such as environmental monitoring,mineral exploration,agriculture and forestry,etc.With the increasing number of hyperspectral data,it has become impossible to provide a certain amount of reference data for each scene to achieve supervised classification.In order to solve this problem,works published have introduced domain adaptation(DA)technology in transfer learning to achieve hyperspectral classification,i.e.,using the classification model trained the source scene(source domain)that is related to the scene to be studied(target domain)to adapt to the target scene.Further,unsupervised classification techniques are particularly important when the source domain associated with the target domain does not exist and there is no reference data in target domain available.Therefore,this thesis focuses on the following three aspects: classification of small target and background pixels in single-frame infrared image,domain adaption for HSI classification and unsupervised HSI classification.In order to better solve these problems,this thesis systematically summarizes the current methods of the three key techniques,and analyzes the shortcomings of these existing methods.Based on these shortcomings,this thesis proposes five new methods for three problems and they are reasonably verified through a large amount of rich real data.Specifically,the main research work and innovations are as follows:1.The thesis reviewed the current methodology of three key issues.The main methods of classification of small target and background pixels in single-frame infrared image,domain adaption for HSI classification and unsupervised HSI classification are summarized.The main veins of their development are sorted out.Meanwhile,the advantages and shortcomings of the existing methods are analyzed,the challenges of the three technologies are summarized,and the key issues that need to be solved are listed.2.The classification of small target and background pixels in single-frame infrared image are studied in detail.1)The thesis proposes an effective infrared small target extraction technique based on the new local contrast measure.First,the difference of Gaussian filters is used to enhance the target and suppress background pixels.Second,a sliding window operation is performed to obtain a fixed-size partial area of the infrared image,the size of which is set larger than the general small target size.Finally,a new local contrast measure is obtained based on the local block to obtain a final small target enhanced map,and an adaptive threshold is used to extract the target region.Experimental results on two real sequences show that the method can achieve better performance.2)From the perspective of local image segmentation,this thesis proposes a small target extraction technique based on facet kernel and random walks(RW),which mainly includes four main stages.First,since the RW algorithm is suitable for images with weak noise,local statistical and mean filtering are applied to smooth the image and eliminate high-brightness speckle noise.Second,facet kernel filtering is utilized to enhance the target pixels and candidate target pixels are extracted by adaptive thresholding.Thirdly,inspired by the infrared small target property,a new local contrast descriptor(NLCD)based on RW algorithm is designed to achieve background suppression and target enhancement.Specifically,the candidate target pixel having the largest gray value is selected as the center pixel to construct a local region,and the NLCD results of all the partial regions are calculated.The calculated NLCD map is then weighted by the facet kernel filtering results to further enhance the target.Finally,the weighted graph is thresholded to achieve target detection.The experimental results based on three data sets show that the proposed method is superior to the traditional comparison method in target extraction accuracy.3.The domain adaption for HSI classification is studied in detail.1)The thesis proposes a HSI domain adaptation classification method based on high-order tensor alignment(TA).Specifically,two HSIs(source domain and target domain)are segmented into superpixels,and adjacent pixels in the same superpixel around the centering pixel are selected to construct a tensor.Domain adaptation is achieved by subspace alignment and dimensionality reduction between two domains,that is,the alignment matrice are used to achieve alignment of two domain subspaces,and the projection matrice are used to map the original tensor to the invariant subspace of the lower dimension.Then the subspace tensors(core tensors)are obtained.In order to preserve the geometric information of the original tensors,a manifold regularization term is used to constrain the optimization process,and the alignment matrice,projection matrice and core tensors are solved under the framework of Tucker decomposition and alternating optimization strategy.In addition,classification performance can be further improved based on subsequent processing strategies for pure sample extraction of each superpixel.2)The thesis proposes a heterogeneous DA(HDA)method for HSI classification under the condition that the number of labeled samples in two domains is limited.The method is implemented in the form of cross-domain collaborative learning(CDCL)and solved by cluster canonical correlation analysis(C-CCA)and random walks(RW)algorithms.Specifically,the CDCL method is an iterative process of three main components,namely,a pseudo-labeling based on the RW algorithm,a cross-domain learning by C-CCA,and a final classification based on an extended RW(ERW)algorithm.First,the initially labeled target samples are regarded as training set,and the pseudo-labeling is achieved by fusing the segmentation results obtained by the RW algorithm and the ERW algorithm,and then the pseudo-labeling result is used to update the training set and extract the target clusters.Second,C-CCA is used to correlate the source samples(source clusters)and target clusters to achieve cross-domain learning.Then,the training set is used to train the classifier model in the projection correlation subspace,the unlabeled target samples are classified,and the corresponding estimated probability map is obtained.The resulting probability map and training set will be further used to update the training set again.Finally,when the iterative process converges,the ERW classifier uses the final training set and the estimated probability map to obtain the final classification result.4.The unsupervised classification problem of HSI is studied in detail.The deep learning technique for HSI clustering based set-to-set distance and sample-to-sample distance is studied.Deep learning techniques have been used to solve HSI supervised classification problems and achieve the most advanced performance.Based on this,the thesis proposes a deep learning algorithm for HSI clustering.The algorithm tries to learn deep feature based on set-to-set and sample-to-sample distances(LSSD).The technique consists of four main components: i)over-segmentation,ii)generation of set-to-set distances and sample-to-sample distances,iii)learning deep embedding by training Siamese network,and iv)density-based spectral clustering.First,the HSI is subdivided into superpixels by using an entropy rate superpixel(ERS)algorithm.Secondly,the distance between the sets is obtained by considering the samples in each superpixel as an affine hull(AH)model,and the distance between the samples calculated by using a local covariance matrix representation(LCMR).Thirdly,similar/dis-similar sample pairs are extracted from the two distances,which are then input into a multilayer perceptron(MLP)network and the deep features are learned by training a Siamese network with contrastive loss.Finally,density-based spectral clustering is applied to the deep features to obtain clustering results.Experimental results on three real HSIs show that the method can achieve better performance.