Hyperspectral Image Classification With Global–Local Discriminant Analysis and Spatial–Spectral Context

Recently, depth methods have developed rapidly and achieved remarkable achievements in the field of hyperspectral image classification. Deep learning largely depends on high-quality labels. However, for hyperspectral images, obtaining high-quality labels is difficult, expensive and time-consuming. For the problem of the small number of training samples, We propose a Spatial-spectral Siamese network, which randomly selects a number of tasks composed of support sets and query sets on the training set. The network is used to extract spectral and spatial features and fuse them to obtain spatial-spectral features. The fused Spatial-spectral features are trained to optimize the Spatial-spectral Siamese network by contrastive loss, and the similarity loss is used to measure the similarity score between samples to optimize the similarity network. In addition, the cluster labels obtained by clustering are used as supervised information, The performance of network classification is further optimized by monitoring the clustering task and using a large amount of unlabeled sample information. The results of two frequently used hyperspectral data sets show that the proposed method still has good performance and effect in classification accuracy when training samples are very small.

This paper introduces a novel and efficient Graph Convolutional Network (GCN) and Spatial Supporting Modification (SSM) method for classifying Hyperspectral Images (HSI), called Spatial First (SPA-F). The proposed method can utilize spatial information fully based on the assumption that neighboring pixels are more likely to belong to the same category. The method uses spatial information from two levels and consists of the following steps. Firstly, a graph is constructed of all non-background pixels. Secondly, a GCN is trained and tested to perform node classification tasks based on the constructed graph. Thirdly, the spatial supporting pixels for all non-background pixels are located according to their spatial position in the HSI. Finally, the proposed SSM strategy is used to fine-tune the initial classification results obtained in the second step. The superiorities of the proposed method are verified on the Indian Pines and Salinas-A scene datasets even when using a small number of training samples.

In this letter, a relaxed clustering assumption and spatial Laplace-regularizer-based semisupervised hyperspectral image classifier is proposed. Considering the mixed pixels and noise intrinsic in hyperspectral image, we relax the clustering assumption employed in most of the available classifiers so that the similar hyperspectral vectors tend to share the “similar” labels instead of the “same” label, to formulate a modified spectral similarity regularizer. Moreover, the spatial homogeneity assumption is cast on hyperspectral pixels to construct a spatial regularizer, to overcome the salt-and-pepper misclassification of images. The effectiveness of our proposed method is evaluated via experiments on AVIRIS data, and the results show that it exhibits state-of-the-art performance, particularly when there are a small number of training samples.

With the wide deployment of wireless networks, channel state information (CSI)-based joint recognition of location and activity boasts a wide range of application prospects in the fields of smart homes, remote health monitoring, and security alarm. However, the existing algorithms for joint recognition of location and activity can achieve high recognition accuracy only when trained on a large number of samples. To address this problem, this study proposes a novel, small-sample joint “location and activity” recognition algorithm based on residual networks (ResNets) with multiattention layers, referred to as the SSJR algorithm. In the algorithm, a small number of training samples are expanded via subtraction between CSI amplitude data of multiple links and horizontal flipping of images. Furthermore, a peak-seeking algorithm is proposed to decompose CSI data into location and activity recognition components, both of which are input into the constructed ResNets with multiattention layers for joint recognition. The algorithm can not only expand number of samples but also accurately extract and enhance the features of each category of samples. Experimental results indicate that the SSJR algorithm achieves an average recognition accuracy of 96.83% with only five training samples per category, demonstrating both its practicality and feasibility for human-computer interaction applications.

Practical applications of online handwritten character recognition demand robust and highly accurate recognition along with low memory requirements. The Active-DTW~\cite{activedtw} classifier proposed by Sridhar {\it et al}. combines the advantages of generative and discriminative classifiers to address the similarity of between-class samples, while taking into account the variability of writing styles within the same character class. Active-DTW uses Active Shape Models to model the significant writing styles in a memory-efficient manner.However, in order to create accurate models, a large number of training samples is needed up front, which is not desirable or available in many practical applications. In this paper, we propose a supervised adaptation framework for the Active-DTW classifier which allows recognition to begin with a small number of training samples, and adapts the classifier to the new samples presented to the system during recognition. We compare the performance of Active-DTW using the proposed adaptation framework, with a nearest-neighbor classifier using an LVQ-based adaptation scheme, on the online handwritten Tamil character dataset.

The performance of a nearest-neighbor classifier is degraded if only a small number of training samples are used as prototypes. An algorithm is presented for modifying the prototypes so that classification rate can be increased. This algorithm makes use of a two-layer perceptron with one second-order input. Each hidden node of the perceptron represents a prototype and the weights of connections between a hidden node and the input nodes are initially set equal to the feature values of the corresponding prototype. The weights are then changed using a gradient-based algorithm to generate a new prototype. The algorithm has been tested with good results. >

Combining spectral and spatial information can significantly improve the classification performance of hyperspectral image (HSI). Currently, a lot of spectral–spatial HSI classification methods have been proposed. However, the task of HSI classification has remained challenging since the number of training samples is limited in real scenarios. In this article, we propose a novel HSI classification framework with single sample, in which the spectral self-similarity and spatial polygon structure information are fully combined to improve the classification performance. On the one hand, spectral self-similarity is used to expand training samples, which makes it possible to obtain sufficient samples with minimal cost. On the other hand, polygonal partition is introduced to acquire the geometrical structure of land covers in man-made environments. Specifically, the edge information of geometric objects is captured by polygonal partition, which can be utilized to constrain the spatial range of sample expansion and optimize the classification results. Experimental results on three real HSIs illustrate that the proposed method performs very well under small training sample size even when the number of samples is single per class.

Multimodal Earth observation data fusion: Graph-based approach in shared latent space

Class imbalance is a challenging problem in many classification tasks. It induces biased classification results for minority classes that contain less training samples than others. Most existing approaches aim to remedy the imbalanced number of instances among categories by resampling the majority and minority classes accordingly. However, the imbalanced level of difficulty of recognizing different categories is also crucial, especially for distinguishing samples with many classes. For example, in the task of clinical skin disease recognition, several rare diseases have a small number of training samples, but they are easy to diagnose because of their distinct visual properties. On the other hand, some common skin diseases, e.g., eczema, are hard to recognize due to the lack of special symptoms. To address this problem, we propose a self-paced balance learning (SPBL) algorithm in this paper. Specifically, we introduce a comprehensive metric termed the complexity of image category that is a combination of both sample number and recognition difficulty. First, the complexity is initialized using the model of the first pace, where the pace indicates one iteration in the self-paced learning paradigm. We then assign each class a penalty weight that is larger for more complex categories and smaller for easier ones, after which the curriculum is reconstructed by rearranging the training samples. Consequently, the model can iteratively learn discriminative representations via balancing the complexity in each pace. Experimental results on the SD-198 and SD-260 benchmark data sets demonstrate that the proposed SPBL algorithm performs favorably against the state-of-the-art methods. We also demonstrate the effectiveness of the SPBL algorithm's generalization capacity on various tasks, such as indoor scene image recognition and object classification.

In this work, we present tensor-based linear and nonlinear models for\nhyperspectral data classification and analysis. By exploiting principles of\ntensor algebra, we introduce new classification architectures, the weight\nparameters of which satisfies the {\\it rank}-1 canonical decomposition\nproperty. Then, we introduce learning algorithms to train both the linear and\nthe non-linear classifier in a way to i) to minimize the error over the\ntraining samples and ii) the weight coefficients satisfies the {\\it rank}-1\ncanonical decomposition property. The advantages of the proposed classification\nmodel is that i) it reduces the number of parameters required and thus reduces\nthe respective number of training samples required to properly train the model,\nii) it provides a physical interpretation regarding the model coefficients on\nthe classification output and iii) it retains the spatial and spectral\ncoherency of the input samples. To address issues related with linear\nclassification, characterizing by low capacity, since it can produce rules that\nare linear in the input space, we introduce non-linear classification models\nbased on a modification of a feedforward neural network. We call the proposed\narchitecture {\\it rank}-1 Feedfoward Neural Network (FNN), since their weights\nsatisfy the {\\it rank}-1 caconical decomposition property. Appropriate learning\nalgorithms are also proposed to train the network. Experimental results and\ncomparisons with state of the art classification methods, either linear (e.g.,\nSVM) and non-linear (e.g., deep learning) indicates the outperformance of the\nproposed scheme, especially in cases where a small number of training samples\nare available. Furthermore, the proposed tensor-based classfiers are evaluated\nagainst their capabilities in dimensionality reduction.\n

High spatial resolution (1–5 m) remotely sensed datasets are increasingly being used to map land covers over large geographic areas using supervised machine learning algorithms. Although many studies have compared machine learning classification methods, sample selection methods for acquiring training and validation data for machine learning, and cross-validation techniques for tuning classifier parameters are rarely investigated, particularly on large, high spatial resolution datasets. This work, therefore, examines four sample selection methods—simple random, proportional stratified random, disproportional stratified random, and deliberative sampling—as well as three cross-validation tuning approaches—k-fold, leave-one-out, and Monte Carlo methods. In addition, the effect on the accuracy of localizing sample selections to a small geographic subset of the entire area, an approach that is sometimes used to reduce costs associated with training data collection, is investigated. These methods are investigated in the context of support vector machines (SVM) classification and geographic object-based image analysis (GEOBIA), using high spatial resolution National Agricultural Imagery Program (NAIP) orthoimagery and LIDAR-derived rasters, covering a 2,609 km2 regional-scale area in northeastern West Virginia, USA. Stratified-statistical-based sampling methods were found to generate the highest classification accuracy. Using a small number of training samples collected from only a subset of the study area provided a similar level of overall accuracy to a sample of equivalent size collected in a dispersed manner across the entire regional-scale dataset. There were minimal differences in accuracy for the different cross-validation tuning methods. The processing time for Monte Carlo and leave-one-out cross-validation were high, especially with large training sets. For this reason, k-fold cross-validation appears to be a good choice. Classifications trained with samples collected deliberately (i.e., not randomly) were less accurate than classifiers trained from statistical-based samples. This may be due to the high positive spatial autocorrelation in the deliberative training set. Thus, if possible, samples for training should be selected randomly; deliberative samples should be avoided.

In this paper, a graph regularized nonlinear ridge regression (RR) model is proposed for remote sensing data analysis, including hyper-spectral image classification and atmospheric aerosol retrieval. The RR is an efficient linear regression method, especially in handling cases with a small number of training samples or with correlated features. However, large amounts of unlabeled samples exist in remote sensing data analysis. To sufficiently explore the information in unlabeled samples, we propose a graph regularized RR (GRR) method, where the vertices denote labeled or unlabeled samples and the edges represent the similarities among different samples. A natural assumption is that the predict values of neighboring samples are close to each other. To further address the nonlinearly separable problem in remote sensing data caused by the complex acquisition process as well as the impacts of atmospheric and geometric distortions, we extend the proposed GRR into a kernelized nonlinear regression method, namely KGRR. To evaluate the proposed method, we apply it to both classification and regression tasks and compare with representative methods. The experimental results show that KGRR can achieve favorable performance in terms of predictability and computation time.

An antagonistic training algorithm for TFT-LCD module mura defect detection

Fine-grained visual categorization aims to distinguish objects in subordinate classes instead of basic class, and is a challenge visual task due to the high correlation between subordinated classes and large intra-class variation (e.g. different object poses). Although, deep convolutional neural network (DCNN) has brought dramatic success on generic object classification, detection and segmentation with the availability of the large-scale training samples, direct application of DCNN on fine-grained visual categorization, where only decades or at most hundreds of training samples for each subordinate class are available in most public finegrained image datasets, cannot lead to satisfactory classification results due to small number of training samples. This study explores the transfer learning strategy for finegrained dog breed categorization based on the learned CNN models with the large-scale image dataset: ImageNet, and prove promising performance with two DCNN models: AlexNet and VGG-16. Furthermore, we argue that different DCNN architecture may extract the representation of different image aspects due to the previously defined CNN kernel sizes, number and various operations in the model learning procedure, and thus result in different performance for visual categorization. This study proposes to fusion multiple CNN architectures for combining different aspect representations to give more accurate performance. We compressively study the fusion of different layers such as Fc6 and Fc7 in AlexNet and VGG-16, and manifest 2.88% improvement of the fusion architecture over the best performance of the only one DCNN model: VGG-16 from 81.2% to 84.08%.

Emergent behavior in swarm robotic systems is key to obtaining complex behavior by a group of relatively simple agents. The question is how to design the individual behaviors of agents in such a way that the desired global behavior emerges. Different approaches have been proposed to solve this problem: from biologically inspired probabilistic behavioral models to evolutionary techniques. In some situations, however, creating a complex probabilistic model of the behavior or developing a proper setup for an evolutionary process can be challenging. In this paper we propose a new method, based on supervised learning on a relatively small number of training samples. We apply our method to the well-known clustering problem and show that this approach yields the desired global clustering behavior.KeywordsIndividual BehaviorObstacle AvoidanceCluster ProblemEmergent BehaviorBasic BehaviorThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.