Semi-Supervised Tensorial Locally Linear Embedding for Feature Extraction Using PolSAR Data

Graph embedding (GE) frameworks are used for extracting the discriminative features of hyperspectral images (HSIs). However, it is difficult to select a proper neighborhood size for graph construction. To overcome this difficulty, a semisupervised feature extraction (FE) method, called semisupervised collaborative representation graph embedding (SCRGE), is proposed. The proposed algorithm utilizes collaborative representation (CR) to obtain the collaborative coefficients of labeled and unlabeled samples. Then, a semisupervised graph is constructed using the collaborative coefficients of the labeled samples within the same class and the collaborative coefficients of the unlabeled samples, and an interclass graph is constructed using the collaborative coefficients of the labeled samples in different classes. Finally, a projection matrix for FE is obtained by embedding these graphs into a low-dimensional space. SCRGE not only inherits the merits of CR to reveal the collaborative reconstructive properties of data but also enhances intraclass compactness and interclass separability to improve the discriminating power for classification. Experimental results on three real HSIs datasets demonstrate that SCRGE outperforms other state-of-the-art FE methods in terms of classification accuracy.

Urban Scene classification using single source data is massively studied in remote sensing field. However, single source only provides one certain perspective of the complicated urban scene while the fusion of multimodal dataset can provide complementary knowledge. We aim at fusing the spectrum information of the hyperspectral image and the scattering mechanisms of PolSAR data for urban scene classification. For the joint usage of the two data sets, a simple concatenation would lead to extraction of insufficient information and weakens the influence of the lower dimensional data. In this work, the end-to-end convolutional neural network is utilized to automatically learn how to effectively extract features and to fuse the hyperspectral image and the PolSAR data. More specifically, we propose a novel two-stream convolutional network architecture. It creates identical but separated convolutional stream for each data. Subsequently, the two streams are merged with comparable numbers of dimensionality within the fusion layer. This architecture ensures the effectively extraction of informative features from both data for the classification purpose and the fusion of the two data in a balanced manner. Experimental results suggest significantly superior performance of the proposed framework, while comparing to other existing fusion methods. To our knowledge, it is the first time that deep convolutional neural network accomplishes the fusion of hyperspectral image and SAR data.

A new face recognition method is proposed in this paper. The proposed method is based on fuzzy regularized linear discriminant analysis (FR-LDA) and combines the regularized linear discriminant analysis (R-LDA) and the fuzzy set theory. R-LDA is based on a new regularized Fisher's discriminant criterion, which is particularly robust against the small sample size problem compared to the traditional one used in LDA. In the proposed method, we calculate the membership degree matrix by Fuzzy K-nearest neighbor (FKNN) and then incorporate the membership degree into the definition of the between-class and within-class scatter matrices and get the fuzzy between-class and within-class scatter matrices. Experimental results obtained on the FERET database demonstrate that the proposed method improves the classification rate performance.

Imaging mass spectrometry is an innovative technique that combines high-resolution microscopic imaging tools with analytical capabilities of spectrometry. It is a powerful tool to determine the spatial distribution of chemical compounds on complex surfaces, for example, for microscale analysis of cells and tissue in biological samples. The result is a large spectral datacube: a three-dimensional (3D) dataset in which surface position and mass spectral distribution are represented. Analysts try to discover ‘features’: correlations in spectral profiles with a recognizable spatial distribution. Techniques for feature extraction and visualization are developed to improve the exploratory analysis of spectral datacubes. The topic of this work is the design and implementation of feature extraction and visualization techniques in high-resolution imaging spectrometry data. Principal Component Analysis (PCA) is interactively used as a governing approach for feature detection. A wide range of visualization techniques are implemented based on extracted features. The thesis is organized as follows. In Chapter 2 (Spectral analysis: a survey), we provide a brief background survey on spectral analysis. The analysis in the proposed approach is divided into three stages: data acquisition, feature extraction and feature visualization. For each stage, a detailed description of currently applied methods is given. The methods most appropriate for this qualitative approach of analysis are chosen as a specific subset. PCA in combination with a binning function is most suited for extracting features from imaging mass spectral data. Both methods increase the signal-to-noise ratio and reduce the amount of data from imaging mass spectrometry. Chapter 3 (PCA-based feature extraction) compares the quality of three different PCA-based methods for detecting and extracting features from spectral datacubes. We discuss preprocessing of mass spectral data, PCA, additional rotational optimization by VARIMAX, and the PARAFAC method for factor regression. The results are compared quantitatively and qualitatively, together with some performance characteristics. For the quantitative comparison, we used a RMSE metric to compare the methods with ground truth spectra under various noise conditions. For the qualitative comparison, we used three criteria to judge the quality of features in the resulting visualizations. These criteria were applied to interpret the visualizations of features. In Chapter 4 (Feature-based registration), a robust method for automatic featurebased registration is developed. The reduction of uncorrelated noise provided by PCA allows high-resolution imaging mass spectrometry datasets to be automatically aligned and combined for high-resolution analysis of large areas. The results clearly show that the entropy-weighted, mean squared error landscape of chemically matched component images can be used to automatically align high-resolution imaging mass spectrometry datasets. Several spectral datacubes are combined to provide better detection and extraction of features. In Chapter 5 (Feature visualization), a visualization technique is described that utilizes principal components to create transfer functions for volume rendering of a spectral datacube. Two types of spectral datacubes are visualized in 3D by direct volume rendering with these transfer functions to control opacity and highlight extracted features. This enables us to visualize the link between the spectral and spatial characteristics of a feature within the spectral datacube. Applications demonstrate the additional value of these visualizations. Chapter 6 (Feature zooming) presents a technique for spectral and/or spatial zooming of extracted features. This technique is especially useful for spatially extended datasets. The combined spectral datasets are too large in size to be explored and visualized using commonly feature extraction and visualization techniques. Analysts are able to select important features or deselect unimportant features for further analysis on different levels of detail. Moreover, features with unwanted artifacts can be removed to reduce noise. Chapter 7 (High-resolution feature visualization) provides an approach to parametrically visualize features in 3D and at the highest resolution possible. Three parameters control the spectral contribution, the level of detail and the level of density on which an extracted feature is represented. This visualization has feature shapes with well-defined borders and provides more insight into the influences of noise on a mass spectral measurement. It is possible to distinguish different peaks according to their difference in density and spatial position, which would not be possible in a separate spectral or spatial view. An application shows how resulting features are visualized and interpreted. The developed tools generate new possibilities to handle, explore, and visualize the large imaging mass spectrometry datasets. A sensitive, selective, and robust approach for feature extraction enables detection and classification of features in different proteomics applications. Multiple feature shapes with high-resolution characteristics can be compared and examined on different levels of detail. These visualizations can provide more detailed molecular insight in the biochemistry of surfaces and improve classification of peptides and proteins.

Efficient dimensionality reduction can involve generative latent variable models such as probabilistic principal component analysis (PPCA) or independent component analysis (ICA). Such models aim to extract a reduced set of variables (latent variables) from the original ones. In most cases, the learning of these models occur within an unsupervised framework where only unlabeled samples are used. In this paper, we investigate the possibility of estimating an independent factor analysis model (IFA), and thus projecting original data onto a lower dimensional space, when prior knowledge on the cluster membership of some training samples is incorporated. We propose to allow this model to learn within a semi-supervised framework in which likelihood of both labeled and unlabeled samples is maximized by a generalized expectation-maximization (GEM) algorithm. Experimental results with real data sets are provided to demonstrate the ability of our approach to find a low dimensional manifold with good explanatory power.

Motor bearings play an important role in industrial production. Regular inspection and fault diagnosis can provide reliable guarantee for the normal operation of the motor. If the potential or early bearing failures are not detected in time, they will bring safety hazards to production operations. In this thesis, distance measure method for feature extraction and KNN (k-nearest neighbor) algorithm for recognition are combined for the purpose of diagnosis of motor bearing. Through experimental verification, the medium KNN algorithm has the most obvious improvement, and the accuracy has increased from 39. 1% to 99. 8% with the help of the proposed feature extraction. Using the KNN algorithm, the effect of time domain feature extraction, frequency domain feature extraction and distance feature extraction are verified. After comparison, it is concluded that the distance feature extraction method can effectively improve the accuracy rate and speed up the training speed of the model.

We propose a semisupervised sparse graph learning (SSGL) method for feature extraction of hyperspectral remote sensing imagery in this paper. The proposed SSGL method aims to build a semisupervised sparse graph that can maximize the class discrimination and preserve the local neighborhood information by combining labeled and unlabeled samples. In our semisupervised sparse graph, we connect labeled samples according to their label information and sparse representation, connect unlabeled samples by sparse representation which set unlabeled training samples set as dictionary. Moreover, by setting sparse connections between labeled sample and unlabeled samples, the label information can be well propagated from labeled samples to unlabeled samples. In this way, the similarity of data points can be well modelled in the spectral feature space, and the data distribution characteristics can be better expressed. Experimental results on real hyperspectral images (HSIs) demonstrate the advantages of our method compared to some related feature extraction methods.

Security agencies face an emerging challenge of identifying and counter the malicious contents spread on the social media by the terrorists. However, text classification techniques are limited by visualization, pre-processing, features extraction, and larger features space. Additionally, change in criminal content require the learning models to identify altered malicious textual contents which poses extra challenge. This study proposes simplified yet adaptable framework that uses a novel features extraction algorithm for extracting features from the textual part of social media contents. The feature extraction considers selective features from only 8 dimensions and follows a six step process. The extracted features are suitably used to train the state vector machine for the classification of the malicious content. The performance of the proposed method is evaluated against other popular feature selection/extraction algorithms like term frequency-inverse document frequency, Gini Index (GI), Chi square statistics, and PCA. Additionally, machine learning classifiers like decision tree, random forest, and Naïve Bayes are also used for classification. Results suggest that the proposed approach consumes less energy on text visualization, pre-processing, and dimensionality reduction. It also reduces the time-space complexity of the features extraction process and is capable to steer according to the changing strategies of the active criminal groups. In addition, it can effectively analyze the propaganda material published by the extremists. It automatically identifies the radical text on social media platforms allowing understanding of the behaviors, characteristics and subsequent blockage of such content.

Fatalities due to road accidents are mainly caused by distracted driving. Driving demands continuous attention of the driver. Certain levels of distraction while driving can cause the driver lose his/her attention which might lead to a fatal accident. Thus, early detection of distraction will help reduce the number of accidents. Several researches have been conducted for automatic detection of driver distraction. Many previous approaches have employed camera based techniques. However, these methods might detect the distraction rather late to warn the drivers. Although neurophysiological signals using Electroencephalography (EEG) have shown to be another reliable indicator of distraction, EEG signals are very complex and the technology is intrusive to the drivers, which creates serious doubt for its implementation. In this study we investigate Galvanic Skin Responses (GSR) using a wrist band wearable and conduct an empirical characterization of driver GSR signals during a naturalistic driving experiment. We explored time and frequency domain to extract relevant features to capture the changes/patterns at the physiological level. Due to the fact that feature extraction is a manual process and to normalize the feature space toward the identification task, we then transform the feature space using linear discriminant dimensionality reduction to discover discriminative bases of the GSR multivariate feature space that identify distraction. That would eliminate both the computational complexity and the redundancies in the manually generated feature space. Due to multi-class nature of the identification task, there might be biases between the distracted and non-distracted categories that can bias the estimation of between- and within-class scatter matrices. Therefore, we incorporated a class dependent weight to calculate the within class scatter matrices. The proposed weight aims to increase the flexibility of the discriminative bases vectors to capture the factors that focus on eliminating the overlap between distracted versus non-distracted in the generalization phase. Our experimental results demonstrated high cross validation accuracies of distraction identification using GSR signals (i.e. 85.19%). Conducting dimensionality reduction using LDA resulted in slight improvement in accuracy (i.e. 85.94%) using only two discriminant bases. The generalization accuracy was further improved by applying our proposed weighting mechanism (i.e. 88.92%).

Sex determination of silkworm pupae is important for silkworm industry. Multivariate analysis methods have been widely applied in hyperspectral imaging spectroscopy for classification. However, these methods require essential steps containing spectra preprocessing or feature extraction, which were not easy determined. Convolutional neural networks (CNNs), which have been employed in image recognition, could effectively learn interpretable presentations of the sample without the need of ad-hoc preprocessing steps. The species of silkworm pupae are usually up to hundreds. Conventional classifiers based on one species of silkworm pupae could not give high performance when explored to other species that not participating in the model building, resulting in bad generalization ability. In this study, a CNN model was trained to automatically identify the sex of silkworm pupae from different years and species based on the hyperspectral spectra. The results were compared with the frequently used conventional machine classifiers including support vector machine (SVM) and K nearest neighbors (KNN). The results showed that CNN outperformed SVM and KNN in terms of accuracy when applied to the raw spectra with 98.03%. However, the performance of CNN decreased to 95.09% when combined with the preprocessed data. Then principal component analysis (PCA) was adopted to reduce data dimensionality and extract features. CNN gave higher accuracy than SVM and KNN based on PCA. The discussion section revealed that CNN had high generalization ability that could classify silkworm pupae from different species with a rather well performance. It demonstrated that HSI technology in combination with CNN was useful in determining the sex of silkworm pupae.

A new face recognition method via semi-discrete decomposition for one sample problem

This paper aims to introduce a novel approach that includes three steps, namely Energy efficient clustering, Disease diagnosis, and an Alert system. Initially, energy-efficient clustering of nodes was conducted, and to render the clustering more optimal, its centroid was optimally selected by a new hybrid algorithm. In addition, this cluster formation was conducted based on constraints such as distance and energy. Further, the disease diagnosis in IoT was performed under two phases namely, “feature extraction and classification”. During feature extraction, the statistical and higher-order features were extracted. These extracted features were then classified via Optimized Deep Convolutional Neural Network (DCNN). To make the classification more precise, the weights of the DCNN were optimally tuned by a new hybrid algorithm referred to as Hybrid Elephant and Moth Flame with Adaptive Learning (HEM-AL). Finally, an alert system was enabled via proposed severity level estimation, which determined the severity of the disease, suggesting patients to visit the hospital. Lastly, the supremacy of the developed approach was examined via evaluation over the other extant techniques. Accordingly, the proposed model attained an accuracy of 0.99 for test case 1, and was 7.41%, 17.34%, and 13.41% better than traditional NN, CNN, and DCNN models.

Much of the real-world image data is unlabeled or mislabeled. Therefore, even if there is no label, if similar images can be grouped together with image data itself and used the group as a label, more image data can be effectively used in various tasks. This will be especially effective when dividing images belonging to the same domain into sub-groups. Therefore, in this study, we propose an image feature extraction method to be used for image clustering. The proposed feature extraction model is the Multi-head Convolutional Autoencoder (MCAE), which is a model composed of multiple encoders in parallel based on the Convolutional Autoencoder (CAE). The proposed model showed about 14% lower test reconstruction loss compared to CAE, and the correlation coefficient between extracted features was about 56% lower. In addition, as the results of clustering based on the extracted features, MCAE-based clustering showed about 3.5 times higher silhouette score than that CAE-based clustering.

The goal of this paper is to present a new fuzzy logic-based feature extraction method, called Fuzzy Generalized Two-Dimensional Fisher's Linear Discriminant (FG-2DFLD) method for face recognition. More specifically, the FG-2DFLD method is an extension of the G-2DFLD method. In FG-2DFLD method, we introduced fuzzy membership values to all the training samples and incorporated them into the within-class and between-class scatter matrices along the row and column directions. The class-wise mean images are also generated by considering the membership values. The fuzzy k-nearest neighbour (FKNN) algorithm is applied for generating the corresponding degrees of class membership. Finally, the feature matrix generated by solving the eigenvalue problem of these scatter matrices actually provide better discrimination information as supported by the simulation results on the AT&T (formally known as ORL) and UMIST face databases using a radial basis function (RBF) neural network.

A novel semi-supervised hyperspectral image classification approach based on spatial neighborhood information and classifier combination