Face Recognition based on Higher-Order Statistics

In this paper, it is proposed a method to detect structural faults or damages using Higher-Order Statistics (HOS). For this, vibration signals were taken from cantilever beams. Such vibrations were generated by a DC motor with varying rotation, generating vibrations at various frequencies. Vibration signals and engine speed control were performed by an Arduino board. After the signal acquisition, parameters are extracted by means of second-, third- and fourthorder cumulants and then the most relevant ones were selected by the Fisher’s Discriminant Ratio (FDR). To fault detection, a Support Vector Machine (SVM) classifier has been designed in its One-Class version, where only oneclass knowledge is required. The results showed a good ability to represent vibration signals via HOS along with a large reduction in dimensionality given using FDR and a good generalization by means of the SVM classifier. Failure detection results showed 100% success rates.

The Fisher's discriminant ratio has been used as a class separability criterion and implemented in a k-means clustering algorithm for performing simultaneous feature selection and data set trimming on a set of 221 HIV-1 protease inhibitors. The total number of molecular descriptors computed for each inhibitor is 43, and they are scaled to lie between 1 and 0 before being subjected to the feature selection process. Since the purpose is to select some of the most class sensitive descriptors, several feature evaluation indices such as the Shannon entropy, the linear regression of selected descriptors on the pKi of selected inhibitors, and a stepwise variable selection program are used to filter them. While the Shannon entropy provides the information content for each descriptor computed, more class sensitive descriptors are searched by both the linear regression and stepwise variable selection procedures. The inhibitors are divided into several different numbers of classes. They are subsequently divided into five classes due to the fact that the best feature selection result is obtained by the division. Most of the good features selected are the topological descriptors, and they are correlated well with the pKi values. The outliers or the inhibitors with less class-sensitive descriptor values computed for each selected descriptor are identified and gathered by the k-means clustering algorithm. These are the trimmed inhibitors, while the remaining ones are retained or selected. We find that 44% or 98 inhibitors can be retained when the number of good descriptors selected for clustering is three. The descriptor values of these selected inhibitors are far more class sensitive than the original ones as evidenced by substantial increasing in statistical significance when they are subjected to both the SYBYL CoMFA PLS and Cerius2 PLS regression analyses.

Schizophrenia is a serious psychiatric illness which needs early and accurate diagnosis. Difference in activation patterns of schizophrenia patients and healthy subjects can be identified with the help of functional magnetic resonance imaging (fMRI). However, manual diagnosis using fMRI depends on subjective observation and may be erroneous. This has motivated the pattern recognition and machine learning research community to develop a reliable and efficient decision model for classification of schizophrenia patients and healthy subjects. However, high dimensionality and low availability of fMRI data leads to the curse-of-dimensionality problem which may degrade the performance of decision model. In the present research work, a combination of feature extraction and feature selection techniques is utilised to obtain a reduced set of relevant and non-redundant features for classification of schizophrenia patients and healthy subjects. Features are extracted from pre-processed fMRI data using a general linear model approach. Next Fisher's discriminant ratio, a univariate method, is employed for feature selection i.e. To determine a subset of discriminative features. Further, minimum Redundancy Maximum Relevance (mRMR) feature selection, a multivariate method, is employed to obtain a set of relevant and non-redundant features which are used for learning a decision model using support vector machine. Two balanced and well-age matched datasets of auditory oddball task derived from a publicly available multisite FBIRN dataset are used for experiments. First dataset consists of fMRI scans of 34 schizophrenia patients and 34 healthy subjects acquired through 1.5 Tesla scanners while second dataset consists of 25 schizophrenia patients and 25 healthy subjects acquired through 3 Tesla scanners. The performance is evaluated in terms of sensitivity, specificity and classification accuracy, and compared with two well-known existing approaches for classification of schizophrenia patients and healthy subjects using fMRI. Experimental results demonstrate that the proposed model outperforms the existing approaches in terms of sensitivity, specificity and classification accuracy. The proposed approach achieves classification accuracy of 88.2% and 78.0% for 1.5 Tesla and 3 Tesla datasets respectively. In addition, the brain regions containing the discriminative features are identified which may be potential biomarkers for diagnosis of schizophrenia using fMRI.

This paper proposes a novel feature ranking method, DensityRank, based on kernel estimation on the feature spaces to improve the classification performance. As the availability of raw data in many of today's applications continues to grow at an explosive rate, it is critical to assess the learning capabilities of different features and select the important subset of features to improve learning accuracy as well as reduce computational cost. In our approach, kernel methods are used to estimate the probability density function for each feature across different class labels. Discrepancy analysis based on the mean integrated square error (MISE) between pairs of such density estimations is used to provide the ranking values. Then, the ranked subspace method is adopted to select subsets of important features that are used to develop the learning models. Comparative study of this method with those of traditional ranking methods related to Fisher's discrimination ratio and information gain theory, as well as the random subspace algorithm and the bootstrap aggregating (bagging), are presented in this paper. Simulation results on various real-world data sets illustrate the effectiveness of the proposed method.

Patient Specific Machine Learning Models for ECG Signal Classification

Tool condition monitoring (TCM) is extremely important to ensure production efficiency and workpiece quality. It is crucial to extract and select suitable features from raw signals to improve the robustness and feasibility of TCM systems. This paper introduces a cutting force singularity analysis approach for TCM in milling, which correlates tool wear states with force waveform variations. The Holder exponents (HEs) were selected as the index of singularities. HEs are calculated by wavelet transform modulus maxima (WTMM). The raw signal is de-noised based on WTMM estimation, which can effectively preserve singularities compared with traditional filters. Fisher’s discriminant ratio (FDR) is employed to rank the discriminant capability of statistical features of HEs. It is found that means of HEs and quantities of singular points estimated from feed forces Fx show the strongest class-discriminant ability. Then, these features were chosen as training samples to propose a TCM approach based on the support vector machine (SVM). Experimental results indicate that this approach provides reliable and effective advice for tool change.

Among numerous artificial intelligence approaches, k-Nearest Neighbor algorithms, genetic algorithms, and artificial neural networks are considered as the most common and effective methods in classification problems in numerous studies. In the present study, the results of the implementation of a novel hybrid feature selection-classification model using the above mentioned methods are presented. The purpose is benefitting from the synergies obtained from combining these technologies for the development of classification models. Such a combination creates an opportunity to invest in the strength of each algorithm, and is an approach to make up for their deficiencies. To develop proposed model, with the aim of obtaining the best array of features, first, feature ranking techniques such as the Fisher's discriminant ratio and class separability criteria were used to prioritize features. Second, the obtained results that included arrays of the top-ranked features were used as the initial population of a genetic algorithm to produce optimum arrays of features. Third, using a modified k-Nearest Neighbor method as well as an improved method of backpropagation neural networks, the classification process was advanced based on optimum arrays of the features selected by genetic algorithms. The performance of the proposed model was compared with thirteen well-known classification models based on seven datasets. Furthermore, the statistical analysis was performed using the Friedman test followed by post-hoc tests. The experimental findings indicated that the novel proposed hybrid model resulted in significantly better classification performance compared with all 13 classification methods. Finally, the performance results of the proposed model was benchmarked against the best ones reported as the state-of-the-art classifiers in terms of classification accuracy for the same data sets. The substantial findings of the comprehensive comparative study revealed that performance of the proposed model in terms of classification accuracy is desirable, promising, and competitive to the existing state-of-the-art classification models.

To improve the efficiency of traditional face recognition techniques, this paper proposes a novel face recognition algorithm called Image Gradient Feature Compensation (IGFC). Based on the gradients along four directions in an image, a fusion algorithm and a compensation method are implemented to obtain features of the original image. In this study, gradient magnitude maps of a face image are calculated along four directions. Fusion gradients and differential fusion gradients are produced by fusing the four gradient magnitude maps of a face image in multiple ways, and they are used as compensation variables to compensate the appropriate coefficients on the original image and build IGFC feature maps of the original face image. Subsequently, IGFC feature maps are divided into several blocks to calculate the concatenated histogram over all blocks, which is in turn utilized as the feature descriptor for face recognition. Principal component analysis (PCA) is used to cut down the number of dimensions in high-dimensional features, which are recognized by the Support Vector Machine (SVM) classifier. Finally, the proposed IGFC method is superior to traditional methods as suggested by verification studies on YALE, ORL, CMU_PIE, and FERET face databases. When the LibSVM parameter was set to ‘-s 0 -t 2 -c 16 -g 0.0009765625’, the algorithm achieved 100% recognition on Yale and ORL data sets, 92.16% on CMU_PIE data sets, and 74.3% on FERET data sets. The average time for simultaneous completion of the data sets examined was 1.93 s, and the algorithm demonstrated a 70.71% higher running efficiency than the CLBP algorithm. Therefore, the proposed algorithm is highly efficient in feature recognition with excellent computational efficiency.

Eye center localization in a facial image based on geometric shapes of iris and eyelid under natural variability

All over the world, people are wearing face masks and practising social distancing to protect themselves against the Coronavirus disease (COVID-19). The need for contactless biometric systems has increased to avoid the common point of contact. Among contactless biometric systems, facial recognition systems are the most economical and effective ones. Conventional face recognition systems rely heavily upon the facial features of the eyes, nose, and mouth. But due to people wearing face masks, the important facial features of the nose and mouth get hidden under the mask, resulting in degraded performance by the facial recognition systems on masked faces. In this paper, we propose a Dense Residual Unit (DRU) aided with Quadruplet loss on top of existing facial recognition systems. This solution tries to unveil the masked faces by producing embeddings for masked faces, which are similar to embeddings of unmasked faces of the same identity but different from embeddings of different identities. We have evaluated our method using two pre-trained facial recognition models’ backbones, i.e. Resnet-101 and MobileFaceNet, and upon two datasets, among them, one is a real-world dataset, i.e. MFR2, and one is a simulated masked face dataset i.e. masked version of LFW. We have achieved improvement in the performance of masked face recognition in terms of False Match Rate, False Non-Match Rate, Fisher Discriminant ratio, and Equal Error Rate.KeywordsBiometricMasked face recognitionMasked face verificationQuadruplet loss

The problem of ultra-high dimensionality small samples exists in the face data. In order to obtain discriminative features with higher generalization ability and improve the face recognition rate, a weighted maximum margin criterion (WMMC) method is presented for extracting discriminative features of face data. The method considers that the actual inter-class scatter is less than the inter-class scatter reflected by the inter-class scatter matrix calculated with limited training samples, and the actual intra-class scatter is more than the intra-class scatter reflected by the intra-class scatter matrix calculated with limited training samples, the WMMC function is obtained by recalculating the inter-class scatter matrix and intra-class scatter matrix. Experimental results using minimum distance (MD) and k-nearest neighbor (kNN) classifiers on ORL and Yale face datasets show that WMMC extracts discriminative features with higher generalization ability and significantly improves face recognition rate compared to maximum margin criterion (MMC).KeywordsFeature extractionWeighted maximum margin criterion (WMMC)Linear discriminant analysis (LDA)Face recognition

Human cognitive state classification using Electroencephalogram (EEG) signal is one of the dynamic exploring regions in emerging smart machine systems. It motivates to build estimator for revealing the human cognitive behaviours for different environments. In this work, we present an operational human cognitive behaviour detection system based on ambulatory EEG signal analysis. A novel event driven environment is created using external stimuli to collect those signals. They are captured using 14 channel Emotiv neuro-headset. Mel-Frequency Cepstral Coefficients (MFCC) is a very popular feature extraction method in acoustic signal processing. In spite of it’s popularity, it has not been explored much when it comes to EEG signals. This paper proposes a novel technique of applying MFCC feature extraction method for ambulatory EEG signals and the results are found to be promising. Besides other feature extraction methods such as: Power Spectral Density (PSD), Discrete Wavelet Transformation (DWT) are used to extract intrinsic features from EEG signal. A statistical technique, Fisher Discriminant Ratio (FDR) is applied on them to evaluate the robustness of each feature. A higher FDR value is observed from MFCC features which demonstrate its discriminative power during classification. Different machine learning techniques such as Probabilistic neural network (P-NN), k-nearest neighbor (kNN), multi-class support vector machine (MCSVM) are applied for training, testing and validating the extracted features. A comparative analysis is discussed for each classifier coupled with each feature. It is found that P-NN with MFCC feature produces the maximum accuracy of 92.61% amongst all the remaining classifiers.

K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) are the two Supervised Machine Learning Algorithms which have been used extensively in solving classification and regression problems in various application domains—medical diagnosis, pattern recognition, and image classification being the popular ones. KNN has been attractive for its simplicity—SVM has been considered as complex. Computational complexity of KNN is less compared to SVM. However, accuracy of SVM in general has been higher compared to KNN. This paper carries out an analysis of complexity of KNN and SVM and then looks at the alternative strategies to reduce the complexity. Accuracies of both SVM and KNN are analyzed from literature survey of experimental results with different training datasets. Reduction of number of training instances and usage of kernel functions are found to be the solutions for reducing SVM complexity. Factors impacting KNN accuracy are analyzed and possible alternatives are assessed. Kernel functions and distance-weighted KNN are concluded as the solution to improve KNN accuracy to compete with SVM accuracy. To utilize the best potential of the two algorithms, the option of a hybrid algorithm combining both KNN and SVM is also proposed.

After decades of extensive research, the field of 2D face recognition is fruitful. However, 2D face recognition is sensitive to changes in pose, facial expression and illumination. 3D face recognition has great potential due to the inherent invariance of pose and illumination changes, but these 3D face images are still difficult to acquire due to several issues such as cost and accessibility. In contrast, RGB-D images captured by low-cost depth sensors such as Kinect are relatively easy to acquire, and research on such low-quality inputs is still limited. In this thesis, an end-to-end multimodal fusion method based on spatial and channel attention is proposed to effectively fuse two image modalities, RGB and depth, to enhance RGB-D face recognition. Our work is specified as follows: firstly, three branches combined with the spatial and channel attention modules are established on the basis of ResNet18, so that the features of three modalities (RGB, depth map, and their fusion modalities, respectively) are obtained. The features under these three modalities are then fused, fed into a shared layer, and the fused features are further learned for deeper discriminative features and passed through a spatial attention vectorization module at the end. The method achieves good results on the Intellifusion RGBD dataset, while we perform ablation trials, which well reveal the contribution of each component of our method to the final performance.

The aim of the present study was to test whether the multifractal properties of ultra-weak photon emission (UPE) from germinating wheat seedlings (Triticum aestivum) change when the seedlings are treated with different concentrations of the toxin potassium dichromate (PD). To this end, UPE was measured (50 seedlings in one Petri dish, duration: approx. 16.6- 28 h) from samples of three groups: (i) control (group C, N = 9), (ii) treated with 25 ppm of PD (group G25, N = 32), and (iii) treated with 150 ppm of PD (group G150, N = 23). For the multifractal analysis, the following steps where performed: (i) each UPE time series was trimmed to a final length of 1000 min; (ii) each UPE time series was filtered, linear detrended and normalized; (iii) the multifractal spectrum (f(α)) was calculated for every UPE time series using the backward multifractal detrended moving average (MFDMA) method; (iv) each multifractal spectrum was characterized by calculating the mode (αmode) of the spectrum and the degree of multifractality (Δα); (v) for every UPE time series its mean, skewness and kurtosis were also calculated; finally (vi) all obtained parameters where analyzed to determine their ability to differentiate between the three groups. This was based on Fisher's discriminant ratio (FDR), which was calculated for each parameter combination. Additionally, a non-parametric test was used to test whether the parameter values are significantly different or not. The analysis showed that when comparing all the three groups, FDR had the highest values for the multifractal parameters (αmode, Δα). Furthermore, the differences in these parameters between the groups were statistically significant (p < 0.05). The classical parameters (mean, skewness and kurtosis) had lower FDR values than the multifractal parameters in all cases and showed no significant difference between the groups (except for the skewness between group C and G150). In conclusion, multifractal analysis enables changes in UPE time series to be detected even when they are hidden for normal linear signal analysis methods. The analysis of changes in the multifractal properties might be a basis to design a classification system enabling the intoxication of cell cultures to be quantified based on UPE measurements.