Search

Hepatocellular Carcinoma (HCC) is the most common malignant liver tumor, being present in 70% of liver cancer cases. It usually evolves on the top of the cirrhotic parenchyma. The most reliable method for HCC diagnosis is the needle biopsy, which is an invasive, dangerous method. In our research, specific techniques for non-invasive, computerized HCC diagnosis are developed, by exploiting the information from ultrasound images. In this work, the possibility of performing the automatic diagnosis of HCC within B-mode ultrasound and Contrast-Enhanced Ultrasound (CEUS) images, using advanced machine learning methods based on Convolutional Neural Networks (CNN), was assessed. The recognition performance was evaluated separately on B-mode ultrasound images and on CEUS images, respectively, as well as on combined B-mode ultrasound and CEUS images. For this purpose, we considered the possibility of combining the input images directly, performing feature level fusion, then providing the resulted data at the entrances of representative CNN classifiers. In addition, several multimodal combined classifiers were experimented, resulted by the fusion, at classifier, respectively, at the decision levels of two different branches based on the same CNN architecture, as well as on different CNN architectures. Various combination methods, and also the dimensionality reduction method of Kernel Principal Component Analysis (KPCA), were involved in this process. These results were compared with those obtained on the same dataset, when employing advanced texture analysis techniques in conjunction with conventional classification methods and also with equivalent state-of-the-art approaches. An accuracy above 97% was achieved when our new methodology was applied.

The acidity of green plum has an important influence on the fruit’s deep processing. Traditional physical and chemical analysis methods for green plum acidity detection are destructive, time-consuming, and unable to achieve online detection. In response, a rapid and non-destructive detection method based on hyperspectral imaging technology was studied in this paper. Research on prediction performance comparisons between supervised learning methods and unsupervised learning methods is currently popular. To further improve the accuracy of component prediction, a new hyperspectral imaging system was developed, and the kernel principle component analysis—linear discriminant analysis—extreme gradient boosting algorithm (KPCA-LDA-XGB) model was proposed to predict the acidity of green plum. The KPCA-LDA-XGB model is a supervised learning model combined with the extreme gradient boosting algorithm (XGBoost), kernel principal component analysis (KPCA), and linear discriminant analysis (LDA). The experimental results proved that the KPCA-LDA-XGB model offers good acidity predictions for green plum, with a correlation coefficient (R) of 0.829 and a root mean squared error (RMSE) of 0.107 for the prediction set. Compared with the basic XGBoost model, the KPCA-LDA-XGB model showed a 79.4% increase in R and a 31.2% decrease in RMSE. The use of linear, radial basis function (RBF), and polynomial (Poly) kernel functions were also compared and analyzed in this paper to further optimize the KPCA-LDA-XGB model.

This study considers the problem of distinguishing between process and sensor faults in nonlinear chemical processes. An integrated fault diagnosis framework is proposed to distinguish chemical process sensor faults from process faults. The key idea of the framework is to embed the cycle temporal algorithm into the dynamic kernel principal component analysis to improve the fault detection speed and accuracy. It is combined with the fault diagnosis method based on the reconstruction-based contribution graph to diagnose the fault variables and then distinguish the two fault types according to their characteristics. Finally, the integrated fault diagnosis framework is applied to the Tennessee Eastman process and acid gas absorption process, and its effectiveness is proved.

Variety classification is an important step in seed quality testing. This study introduces t-distributed stochastic neighbourhood embedding (t-SNE), a manifold learning algorithm, into the field of hyperspectral imaging (HSI) and proposes a method for classifying seed varieties. Images of 800 maize kernels of eight varieties (100 kernels per variety, 50 kernels for each side of the seed) were imaged in the visible- near infrared (386.7–1016.7 nm) wavelength range. The images were pre-processed by Procrustes analysis (PA) to improve the classification accuracy, and then these data were reduced to low-dimensional space using t-SNE. Finally, Fisher’s discriminant analysis (FDA) was used for classification of the low-dimensional data. To compare the effect of t-SNE, principal component analysis (PCA), kernel principal component analysis (KPCA) and locally linear embedding (LLE) were used as comparative methods in this study, and the results demonstrated that the t-SNE model with PA pre-processing has obtained better classification results. The highest classification accuracy of the t-SNE model was up to 97.5%, which was much more satisfactory than the results of the other models (up to 75% for PCA, 85% for KPCA, 76.25% for LLE). The overall results indicated that the t-SNE model with PA pre-processing can be used for variety classification of waxy maize seeds and be considered as a new method for hyperspectral image analysis.

Smartphones are increasingly becoming an efficient platform for solving indoor positioning problems. Fingerprint-based positioning methods are popular because of the wide deployment of wireless local area networks in indoor environments and the lack of model propagation paths. However, Wi-Fi fingerprint information is singular, and its positioning accuracy is typically 2–10 m, thus, it struggles to meet the requirements of high-precision indoor positioning. Therefore, this paper proposes a positioning algorithm that combines Wi-Fi fingerprints and visual information to generate fingerprints. The algorithm involves two steps: merged-fingerprint generation and fingerprint positioning. In the merged-fingerprint generation stage, the kernel principal component analysis feature of the Wi-Fi fingerprint and the local binary pattern features of the scene image are fused. In the fingerprint positioning stage, a light gradient boosting machine (LightGBM) is trained with mutually exclusive feature bundling and histogram optimization to obtain an accurate positioning model. The method is tested in an actual environment. The experimental results show that the positioning accuracy of the LightGBM method is 90% within a range of 1.53 m. Compared with the single-fingerprint positioning method, the accuracy is improved by more than 20%, and the performance is improved by more than 15% compared with other methods. The average locating error is 0.78 m.

High-resolution synthetic aperture radar (SAR) images are mostly used in the current field of ship classification, but in practical applications, moderate-resolution SAR images that can offer wider swath are more suitable for maritime surveillance. The ship targets in moderate-resolution SAR images occupy only a few pixels, and some of them show the shape of bright spots, which brings great difficulty for ship classification. To fully explore the deep-level feature representations of moderate-resolution SAR images and avoid the “dimension disaster”, we innovatively proposed a feature fusion framework based on the classification ability of individual features and the efficiency of overall information representation, called maximum-information-minimum-redundancy (MIMR). First, we applied the Filter method and Kernel Principal Component Analysis (KPCA) method to form two feature subsets representing the best classification ability and the highest information representation efficiency in linear space and nonlinear space. Second, the MIMR feature fusion method is adopted to assign different weights to feature vectors with different physical properties and discriminability. Comprehensive experiments on the open dataset OpenSARShip show that compared with traditional and emerging deep learning methods, the proposed method can effectively fuse non-redundant complementary feature subsets to improve the performance of ship classification in moderate-resolution SAR images.

The construction of prestressed concrete bridges has witnessed a steep increase for the past 50 years worldwide. The constructed bridges exposed to various environmental conditions deteriorate all along their service life. One such degradation is corrosion, which can cause significant damage if it occurs on the main structural components, such as prestressing tendons. In this study, a novel non-destructive evaluation method to incorporate a movable yoke system with denoising algorithm based on kernel principal component analysis is developed and applied to identify the loss of cross-sectional area in corroded external prestressing tendons. The proposed method using denoised output voltage signals obtained from the measuring device appears to be a reliable and precise monitoring system to detect corrosion with less than 3% sectional loss.

This paper presents a comparative study on the performance of different sizes of sensor sets on polymer electrolyte membrane (PEM) fuel cell fault diagnosis. The effectiveness of three sizes of sensor sets, including fuel cell voltage only, all the available sensors, and selected optimal sensors in detecting and isolating fuel cell faults (e.g., cell flooding and membrane dehydration) are investigated using the test data from a PEM fuel cell system. Wavelet packet transform and kernel principal component analysis are employed to reduce the dimensions of the dataset and extract features for state classification. Results demonstrate that the selected optimal sensors can provide the best diagnostic performance, where different fuel cell faults can be detected and isolated with good quality.

A novel classification model, named the quantum-behaved particle swarm optimization (QPSO)-based weighted multiple kernel extreme learning machine (QWMK-ELM), is proposed in this paper. Experimental validation is carried out with two different electronic nose (e-nose) datasets. Being different from the existing multiple kernel extreme learning machine (MK-ELM) algorithms, the combination coefficients of base kernels are regarded as external parameters of single-hidden layer feedforward neural networks (SLFNs). The combination coefficients of base kernels, the model parameters of each base kernel, and the regularization parameter are optimized by QPSO simultaneously before implementing the kernel extreme learning machine (KELM) with the composite kernel function. Four types of common single kernel functions (Gaussian kernel, polynomial kernel, sigmoid kernel, and wavelet kernel) are utilized to constitute different composite kernel functions. Moreover, the method is also compared with other existing classification methods: extreme learning machine (ELM), kernel extreme learning machine (KELM), k-nearest neighbors (KNN), support vector machine (SVM), multi-layer perceptron (MLP), radical basis function neural network (RBFNN), and probabilistic neural network (PNN). The results have demonstrated that the proposed QWMK-ELM outperforms the aforementioned methods, not only in precision, but also in efficiency for gas classification.

Automatic classification of fruits via computer vision is still a complicated task due to the various properties of numerous types of fruits. We propose a novel classification method based on a multi-class kernel support vector machine (kSVM) with the desirable goal of accurate and fast classification of fruits. First, fruit images were acquired by a digital camera, and then the background of each image was removed by a split-and-merge algorithm; Second, the color histogram, texture and shape features of each fruit image were extracted to compose a feature space; Third, principal component analysis (PCA) was used to reduce the dimensions of feature space; Finally, three kinds of multi-class SVMs were constructed, i.e., Winner-Takes-All SVM, Max-Wins-Voting SVM, and Directed Acyclic Graph SVM. Meanwhile, three kinds of kernels were chosen, i.e., linear kernel, Homogeneous Polynomial kernel, and Gaussian Radial Basis kernel; finally, the SVMs were trained using 5-fold stratified cross validation with the reduced feature vectors as input. The experimental results demonstrated that the Max-Wins-Voting SVM with Gaussian Radial Basis kernel achieves the best classification accuracy of 88.2%. For computation time, the Directed Acyclic Graph SVMs performs swiftest.

Facial expression recognition is an interesting and challenging subject. Considering the nonlinear manifold structure of facial images, a new kernel-based manifold learning method, called kernel discriminant isometric mapping (KDIsomap), is proposed. KDIsomap aims to nonlinearly extract the discriminant information by maximizing the interclass scatter while minimizing the intraclass scatter in a reproducing kernel Hilbert space. KDIsomap is used to perform nonlinear dimensionality reduction on the extracted local binary patterns (LBP) facial features, and produce low-dimensional discrimimant embedded data representations with striking performance improvement on facial expression recognition tasks. The nearest neighbor classifier with the Euclidean metric is used for facial expression classification. Facial expression recognition experiments are performed on two popular facial expression databases, i.e., the JAFFE database and the Cohn-Kanade database. Experimental results indicate that KDIsomap obtains the best accuracy of 81.59% on the JAFFE database, and 94.88% on the Cohn-Kanade database. KDIsomap outperforms the other used methods such as principal component analysis (PCA), linear discriminant analysis (LDA), kernel principal component analysis (KPCA), kernel linear discriminant analysis (KLDA) as well as kernel isometric mapping (KIsomap).

Multichannel physiological datasets are usually nonlinear and separable in the field of emotion recognition. Many researchers have applied linear or partial nonlinear processing in feature reduction and classification, but these applications did not work well. Therefore, this paper proposed a comprehensive nonlinear method to solve this problem. On the one hand, as traditional feature reduction may cause the loss of significant amounts of feature information, Kernel Principal Component Analysis (KPCA) based on radial basis function (RBF) was introduced to map the data into a high-dimensional space, extract the nonlinear information of the features, and then reduce the dimension. This method can provide many features carrying information about the structure in the physiological dataset. On the other hand, considering its advantages of predictive power and feature selection from a large number of features, Gradient Boosting Decision Tree (GBDT) was used as a nonlinear ensemble classifier to improve the recognition accuracy. The comprehensive nonlinear processing method had a great performance on our physiological dataset. Classification accuracy of four emotions in 29 participants achieved 93.42%.

Locomotion mode identification is essential for the control of a robotic rehabilitation exoskeletons. This paper proposes an online support vector machine (SVM) optimized by particle swarm optimization (PSO) to identify different locomotion modes to realize a smooth and automatic locomotion transition. A PSO algorithm is used to obtain the optimal parameters of SVM for a better overall performance. Signals measured by the foot pressure sensors integrated in the insoles of wearable shoes and the MEMS-based attitude and heading reference systems (AHRS) attached on the shoes and shanks of leg segments are fused together as the input information of SVM. Based on the chosen window whose size is 200 ms (with sampling frequency of 40 Hz), a three-layer wavelet packet analysis (WPA) is used for feature extraction, after which, the kernel principal component analysis (kPCA) is utilized to reduce the dimension of the feature set to reduce computation cost of the SVM. Since the signals are from two types of different sensors, the normalization is conducted to scale the input into the interval of [0, 1]. Five-fold cross validation is adapted to train the classifier, which prevents the classifier over-fitting. Based on the SVM model obtained offline in MATLAB, an online SVM algorithm is constructed for locomotion mode identification. Experiments are performed for different locomotion modes and experimental results show the effectiveness of the proposed algorithm with an accuracy of 96.00% ± 2.45%. To improve its accuracy, majority vote algorithm (MVA) is used for post-processing, with which the identification accuracy is better than 98.35% ± 1.65%. The proposed algorithm can be extended and employed in the field of robotic rehabilitation and assistance.

Hyperspectral images (HSI) provide rich information which may not be captured by other sensing technologies and therefore gradually find a wide range of applications. However, they also generate a large amount of irrelevant or redundant data for a specific task. This causes a number of issues including significantly increased computation time, complexity and scale of prediction models mapping the data to semantics (e.g., classification), and the need of a large amount of labelled data for training. Particularly, it is generally difficult and expensive for experts to acquire sufficient training samples in many applications. This paper addresses these issues by exploring a number of classical dimension reduction algorithms in machine learning communities for HSI classification. To reduce the size of training dataset, feature selection (e.g., mutual information, minimal redundancy maximal relevance) and feature extraction (e.g., Principal Component Analysis (PCA), Kernel PCA) are adopted to augment a baseline classification method, Support Vector Machine (SVM). The proposed algorithms are evaluated using a real HSI dataset. It is shown that PCA yields the most promising performance in reducing the number of features or spectral bands. It is observed that while significantly reducing the computational complexity, the proposed method can achieve better classification results over the classic SVM on a small training dataset, which makes it suitable for real-time applications or when only limited training data are available. Furthermore, it can also achieve performances similar to the classic SVM on large datasets but with much less computing time.