Emotions are the most powerful information source to study the cognition, behaviour, and medical conditions of a person. Accurate identification of emotions helps in the development of affective computing, brain–computer interface, medical diagnosis system, etc. Electroencephalogram (EEG) signals are one such source to capture and study human emotions. In this Letter, a novel time-order representation based on the S-transform and convolutional neural network (CNN) is proposed for the identification of human emotions. EEG signals are transformed into time-order representation (TOR) based on the S-transform. This TOR is given as an input to CNN to automatically extract and classify the deep features. Emotional states of happiness, fear, sadness, and relax are classified with an accuracy of 94.58%. The superiority of the method is judged by evaluating four performance parameters and comparing it with existing state-of-the-art on the same dataset.

Identification of human emotion involving electroencephalogram (EEG) signals has become an emerging field in health monitoring application as EEG signals can give us a more diverse insight on emotional states. The aim of this study is to develop an efficient framework based on deep learning concept for automatic identification of human emotion from EEG signals. In the proposed framework, the signals are pre-processing for removing noises by low-pass filtering and then delta rhythm is extracted. After that, the extracted rhythm signals are converted into the EEG rhythm images by employing the continuous wavelet transform and then deep features are discovered by using a pre-trained convolutional neural networks model. Afterwards, MobileNetv2 is used for deep feature selection to obtain the most efficient features and finally, long short term memory method is employed for classification of selected features. The proposed methodology is tested on ‘DEAP EEG data set’ (publicly available). This study considers two emotions namely ‘Valence’ and ‘Arousal’ for classification. The experimental results demonstrate that the proposed approach produced accuracies of 96.1% for low/high valence and 99.6% for low/high arousal classification. A further comparison of the proposed method is also carried out and it is seen that the proposed method outperforms other compared methods.

Emotion recognition is actively used in brain–computer interface, health care, security, e-commerce, education and entertainment applications to increase and control human–machine interaction. Therefore, emotions affect people's lives and decision-making mechanisms throughout their lives. However, the fact that emotions vary from person to person, being an abstract concept and being dependent on internal and external factors makes the studies in this field difficult. In recent years, studies based on electroencephalography (EEG) signals, which perform emotion analysis in a more robust and reliable way, have gained momentum. In this article, emotion analysis based on EEG signals was performed to predict positive and negative emotions. The study consists of four parts. In the first part, EEG signals were obtained from the GAMEEMO data set. In the second stage, the spectral entropy values of the EEG signals of all channels were calculated and these values were classified by the bidirectional long-short term memory architecture in the third stage. In the last stage, the performance of the deep-learning architecture was evaluated with accuracy, sensitivity, specificity and receiver operating characteristic (ROC) curve. With the proposed method, an accuracy of 76.91% and a ROC value of 90% were obtained.

Brain complexity and non-stationary nature of electroencephalography (EEG) signal make considerable challenges for the accurate identification of different motor-imagery (MI) tasks in brain–computer interface (BCI). In the proposed Letter, a novel framework is proposed for the automated accurate classification of MI tasks. First, raw EEG signals are denoised with multiscale principal component analysis. Secondly, denoised signals are decomposed by empirical wavelet transform into different modes. Thirdly, the two-dimensional (2D) modelling of modes is introduced to identify the variations of different signals. Fourthly, a single geometrical feature name as, a summation of distance from each point relative to a coordinate centre is extracted from 2D modelling of modes. Finally, the extracted feature vectors are provided to the feedforward neural network and cascade forward neural networks for classification check. The proposed study achieved 95.3% of total classification accuracy with 100% outcome for subject with very small training samples, which is outperforming existing methods on the same database.

The unnatural activities of brain due to seizure events are analysed by electroencephalogram (EEG) signals which are captured from the brain. In this work, a methodology is proposed to classify the seizure EEG signals. In the proposed method, a novel sparse spectrum based empirical wavelet transform (SS-EWT) is applied to decompose the EEG signal into coefficients. From the obtained SS-EWT coefficients, the cross-information potential and normalised energy are extracted as features. Then these features are ranked using the RELIEFF method to obtain significant features. After ranking, these features are fed into the k-nearest neighbour (k-NN) classifier to classify EEG signals corresponding to different brain activities. In this work, the first classification problem is the classification of the seizure (S), seizure-free (F), and normal (Z) EEG signals in which obtained classification accuracy (Acc) is . The second classification problem is the classification of S and Z EEG signals in which Acc is achieved by the proposed method.

Autism is a type of neurodevelopment disorder in which individuals often have difficulties in social relationship, communication, expressing and controlling emotions and poor eye contact, among other symptoms. Currently, electroencephalography (EEG) is the most popular tool to investigate the presence of autism biomarkers. Generally, EEG recordings generate huge volume data with dynamic behaviour. In current practice, the massive EEG data are visually analysed by specialist clinicians to identify autism, which is time consuming, costly, subject to human error, and reduces decision-making reliability. Hence this Letter aims to develop an efficient autism diagnostic system that can automatically identify autism based on time–frequency spectrogram images from EEG signals. Firstly, the raw EEG data is pre-processed using several techniques, such as re-referencing, filtering and normalisation. After that, the pre-processed EEG signals are converted to two-dimensional images using a short-time Fourier transform. Then, textural features are extracted, and significant features are selected using principal component analysis, and feed to support vector machine classifier for classification. The proposed system achieved an average of 95.25% accuracy in ten-fold cross-validation evaluation. The developed system's simplicity and performance indicates usefulness as a decision support tool for healthcare professionals in autism diagnosis.

Electroencephalogram (EEG) measures brainwaves that have been the widely used modality for brain–computer interface (BCI) applications. EEG signal with machine learning has gained substantial success in the BCI. However, the availability of limited training data, appropriate model selection and high false-positive rates are the challenges that need immediate attention. Therefore, in this Letter, the authors present a mental task classification model based on the notion of transfer learning that addresses the issue of data scarcity, model selection and misclassification ratio. In the framework, the proposed model uses pre-trained network for the extraction of diverse feature and classify using support vector machine. The authors employed four pre-trained networks to identify the optimal network for the proposed framework: Vgg16, Vgg19, Resnet18 and Resnet50. The highest classification accuracy of 86.85% (using Resnet50) was achieved using transfer learning. Comparison results showed that convolutional neural network-based approach outperformed conventional machine learning approaches and hence it can be concluded that the EEG-based classification of the mental task using transfer learning model could be used in developing a superior model despite the limited data availability.

Early drowsiness detection may be crucial for the vehicle alertness system. Towards this, wearable technology, camera-based biophysical signals like electroencephalogram (EEG) approaches are utilised. In this Letter, the EEG-based approach is proposed to detect drowsiness. The proposed method consists of random sampling-based artificial signal augmentation, wavelet packet transform decomposition, logarithmic energy entropy, and one-dimensional region mean local binary pattern (1d-RMLBP) based feature extraction and classifier. k-Nearest neighbour and support vector machine classifiers are employed to detect the drowsiness. The MIT/BIH polysomnographic dataset has been used to test the proposed model. The proposed method has superior performance than the other methods using the same data set. The experimental results demonstrate that the proposed model could efficiently detect drowsiness from polysomnographic EEG signals.

Electroencephalogram (EEG) is a diagnostic method that provides information about the functioning of the brain. EEG can be used to diagnose the abnormally functioning part of the brain by monitoring electrical activities. An automated system has been proposed to create a computer-based expert opinion needed in the detection of epilepsy and to capture a more objective view. To this end, the approximation and detail coefficients of EEG signals are calculated by using the wavelet packet transform (WPT). The coefficients were subjected to feature extraction using dispersion entropy and line length methods. The extracted feature vector has been applied as input to the support vector machine (SVM) and k-nearest neighbour (KNN) classifiers. The proposed method was tested using the public EEG seizure dataset created by the University of Bonn. In this study, the dataset was evaluated in two different ways as binary cases and multiclass cases. Evaluated classification accuracy was 100% for binary classification with SVM. For multiclass classification evaluated accuracy was 99.85% with KNN. The proposed method was compared with other methods in the literature using the same dataset. The comparison results provide the superiority of the proposed method.

Drowsiness refers to the state of being sleepy or the state of minimal concentration. It is characterised by a decrease in a person's memory capacity and brain information processing speed. These conditions cause hazards in the real-time working environment such as driving, monitoring power generation and patient health etc. These hazards can be sidestepped by introducing the automatic drowsiness detection system. This Letter suggested the electroencephalogram (EEG)-based automatic drowsiness detection method. The clustering variational mode decomposition (CVMD) explores the non-stationary behaviour of EEG for drowsiness detection. In CVMD optimum allocation sampling analyses non-homogenous EEG signals and converts those into homogeneous EEG clusters. These clusters are then decomposed into band-limited modes. The oscillatory mode characteristics are extracted in terms of several features. These features are fed as input into the least-squares support vector machine classifier. The proposed method provides better drowsiness detection performance in comparison with the different methods using the same data set.

Single modality brain–computer interface (BCI) systems often mislabel the electroencephalography (EEG) signs as a command, even though the participant is not executing some task. In this Letter, the classification of different working memory load levels is presented using a hybrid BCI system. N-back cognitive tasks such as 0-back, 2-back, and 3-back are used to create working memory load on participants while recording EEG and functional near-infrared spectroscopy (fNIRS) signals simultaneously. A combination of statistically significant features obtained from EEG and fNIRS corresponding to frontal region channels are used to classify different N-back commands. Kernel-based support vector machine (SVM) classifiers are employed with and without cross-validation schemes. Classification accuracy of 100% is achieved for binary classification of 0-back against 2-back and 0-back against 3-back using linear SVM, quadratic SVM, and cubic SVM under holdout data division protocol.

The brain–computer interface (BCI) system is doing wonders for people suffering from restricted physical abilities due to accidents or diseases. BCI requires recording of electroencephalogram (EEG) from the subject in order to control electrical devices. In the wireless BCI system, there is always a possibility of EEG signal tampering/attacking during transmission, which may result in malfunction of the BCI system. This work introduces a security block in the BCI system that can ensure that the recorded EEG signals are intact at the receiving end of wireless BCI system. The security block identifies any tampering as well as authenticates the EEG signals at receiving end. Any change in EEG signal can result into classification error; because of that prediction error expansion based reversible watermarking approach has been used. The proposed scheme works efficiently and performance of BCI system remains unaltered after inclusion of security block.

P300 speller-based brain–computer interface (BCI) is an immediate correspondence between the human brain and computer that depends on the translation of mind reactions produced by the stimulus of a subject utilising the P300 speller. No muscle movements are required for this communication. As a P300 paradigm, a novel 2 × 3 matrix consisting of visual home appliances is proposed, which helps disabled people ease their lives by accessing mobile, light, fan, door, television, electric heater etc. In most of the current P300-based BCIs, 5–15 trials work better and the low information transfer rate (ITR) is a major issue in its adaptation in real-time. The objective of this Letter is to improve accuracy as well as an ITR for real-time home appliance control applications. To address this, the authors proposed a single trial weighted ensemble of compact convolution neural network and obtained an ITR of 46.45 bits per minute and an average target appliance accuracy of 93.22% for the BCI-based home environment system. The experimental findings confirmed the feasibility of the proposed method and thus can provide guidance for future use of the system for paralysed patients.

Wilson's disease (WD) is caused by the excessive accumulation of copper in the brain and liver, leading to death if not diagnosed early. WD shows its prevalence as white matter hyperintensity (WMH) in MRI scans. It is challenging and tedious to classify WD against controls when comparing visually, primarily due to subtle differences in WMH. This Letter presents a computer-aided design-based automated classification strategy that uses optimised transfer learning (TL) utilising two novel paradigms known as (i) MobileNet and (ii) the Visual Geometric Group-19 (VGG-19). Further, the authors benchmark TL systems against a machine learning (ML) paradigm. Using four-fold augmentation, VGG-19 is superior to MobileNet demonstrating accuracy and area under the curve (AUC) pairs as 95.46 ± 7.70%, 0.932 (p < 0.0001) and 86.87 ± 2.23%, 0.871 (p < 0.0001), respectively. Further, MobileNet and VGG-19 showed an improvement of 3.4 and 13.5%, respectively, when benchmarked against the ML-based soft classifier – Random Forest.

Multimodal brain MR image analysis is still a challenging research area due to its complex intensity distribution and sensitivity towards the noise. Tumourous cells have different characteristics than normal human cells, which makes them more salient. In this Letter, the authors propose a novel unsupervised spatial information based saliency boosting tumour detection method which will help to identify tumourous cells by making it more clearly visible. Initially, a pseudo-coloured MR image is formed using the CIELab colour space. Saliency map has been established by calculating distance among scales varying elliptical windows in both spatial and colour space. Elliptical windows endeavour to cover-up curved outliers of the brain images. The average intensity value is kept constant by fixing the axis ratio for each window. The proposed algorithm has been evaluated on both real and simulated brain images of different patients from MICCAI-BRATS database. The performance analysis of the new algorithm exhibits higher accuracy with a low computational complexity as compared to other state of the art. The efficacy is due to the immobility of the window across rows and columns to move over the image. The novelty of the proposed technique is that neither it downscales the input images nor require any training bases.

A new design of non-invasive deep bring stimulation system is presented that works based on temporally interfering (TI) electrical fields and high-spatial resolution. The proposed system exploits an increased number of exciting electrical current sources placed around the head transracially to adjust the area and position of the desired region for brain deep neural activation. Only two operational frequencies are used. It is shown that the previously reported TI method can be effectively and systematically extended to focus the amplitude-modulated electric field on the desired location. Results demonstrate that by increasing the number of exciting sources to 40, the spatial resolution of the proposed system can be decreased at least by a factor of 5.8.

Biomedical data are being collected for fields like cancer diagnosis and prognosis, brain signals, speech signals, genetic engineering to name a few. These data are very high dimensional these days, which makes it difficult to extract knowledge out of it through machine learning algorithms. In this work, the authors proposed a hybrid feature selection method based on a multi-attribute decision-making method PROMETHEE (preference ranking organisation method for enrichment evaluations) and Jaya optimisation algorithm. Their proposed method works in two phases. In the first phase, five filter methods are applied to get the ranking for each feature of the data set. In the second phase, all the five individual ranks are used as input choices for PROMETHEE which gives us a final rank for all the features. Then the top 3% features are selected for training the machine learning model. This technique is applicable for feature reduction in any high-dimensional biomedical data. Here, they have studied Parkinson's disease data set. The result shows that the proposed method improves the classification accuracy by 13.73% and that too in a minimum amount of time with a minimum number of features. Hence, this method can be used as an essential pre-processing step for high-dimensional biomedical data.

This Letter illustrates prefrontal haemodynamics as a neurovascular basis of inter-personal working memory differences. A functional near-infrared spectroscopy with sampling frequency ∼2 Hz is used to record the blood oxyhaemoglobin and deoxyhaemoglobin signals from 19 subjects engaged in working memory task of encoding and retrieval of ten symbol-meaning association learning. The individual difference in working memory performance is classified by supervised learning-based linear discriminant analysis and ensemble classifiers. Prior to the classification approach, individual performance is labelled as high, moderate and low on the basis of the performance index. The spontaneous haemodynamic activity and task-evoked responses are marked as background and foreground signals, respectively, which are scaled by means of stream-independent and stream-dependent models. The classifiers' performance shows that the stream-dependent model-based feature construction-classification improves classification accuracy to a major extent compared to the stream-independent model and no gain model. To understand the neurovascular basis of the inter-individual performance difference, diffused voxel plots are constructed. The voxel plots showed that concurrent activation of orbitofrontal and dorsolateral prefrontal cortex could have a possible association with persons' higher working memory performance.

A bandpass frequency selective surface (FSS) with fourth-order bandpass filtering response is proposed based on aperture-coupled patch resonators (ACPRs) under the operation of TM₀₁₀ and TM₀₃₀ resonant modes. An initial bandpass ACPRs FSS with second-order filtering response is firstly realised by periodically arranging the back-to-back patch resonators with coupling apertures etched on the common ground. In order to combine the resonant modes TM₀₁₀ and TM₀₃₀ for a higher-order filtering response, seven pairs of shorting pins are properly loaded based on the cavity model theory for the microstrip patch resonator. The loaded shorting pins make the TM₀₁₀ mode of the ACPRs shift towards the higher frequency, while the TM₀₃₀ mode keeps nearly unchanged. As a result, these two resonant modes can be allocated in proximity to each other to finally realise a fourth-order bandpass filtering response. To validate the design concept, an FSS prototype is designed and fabricated. Reasonable agreement between the measured and simulated results validates the conceptual design.

Attention mechanisms are widely used in re-identification (reID) tasks, but few attention-based architectures have considered integrating local features with their global dependencies, that is the previous works do not model the semantic interdependencies in both spatial dimension and channel dimension. Intuitively, for person feature representations, it is important to model the interdependencies between human body semantics. In this Letter, the authors proposed a dual semantic interdependencies attention module to capture semantic interdependencies in both spatial dimension and channel dimension simultaneously. Besides, they designed a deep supervision branch to directly guide the training of the attention modules and innovatively introduce a channels random dropping mechanism in the training phase to promote the attention modules to capture more discriminative information. Extensive experimental results show that the network merging the above strategies achieves state-of-the-art results on the mainstream reID data sets.

The Letter introduces a novel filtering method based on convolutional neural networks (CNNs) for quality enhancement of light field (LF) images captured by a plenoptic camera and compressed using high-efficiency video coding (HEVC). The method takes advantage of the macro-pixel (MP) structure specific to the LF images and proposes a novel MP-wise filtering approach based on a novel deep neural network architecture. The proposed CNN-based method achieves an outstanding performance when HEVC is employed without its in-loop filters. Experimental results show high luminance-peak signal-to-noise ratio (Y-PSNR) gains and average Y-Bjøntegaard delta (BD)-rate savings of over HEVC on a large data set.

In this Letter, the authors propose a novel mask-guided class activation mapping (MCAM) network for person re-identification, which learns background-invariant and view-invariant features. Specifically, a novel loss function named mask-guided mapping loss is meticulously formulated to utilise the human binary masks, which contain helpful body shape information as the reference standard, thereby guiding the model to place more emphasis on human body regions. Moreover, they propose a new weighted channel attention (WCA) module, which replaces the global average pooling with a global depthwise convolution layer. By virtue of this particular WCA module, the feature information distributed across the spatial space can be individually weighted and dynamically compressed into a more precise channel attention map. Extensive experiments have been carried out on three widely-used re-identification data sets. Compared with the baseline model, MCAM has gained rank-1 accuracy improvement of 2.0% on Market-1501, 6.0% on DukeMTMC-reID, and 7.5% on CUHK03-NP, confirming its effectiveness.

Nowadays, deep learning-based palmprint recognition methods have achieved great success. However, they are mainly focused on the accuracy and ignore the privacy, which is more important in the practical applications. In this Letter, a novel method, federated hash learning (FHL), is proposed for privacy palmprint recognition. There are several agents deploy in different communities, and they have different models and private data. An available public dataset is introduced to provide communications for each agent. Through appropriate federated loss, the agents are connected to help each other train the models to improve the accuracy. Experiments are conducted on constrained and unconstrained palmprint benchmarks. The results demonstrate that the authors’ FHL can outperform other baselines and obtain promising accuracy.

The suppression of leakage current via surface passivation plays a critical role for GaSb-based optoelectronic devices. In this Letter the authors carefully optimise the sulfur passivation parameters for improving the performance of GaSb p–i–n devices. Two competing processes are evaluated during the sulfur passivation process: the hydrolysis of HS^– ions that aide surface passivation and the re-oxidation, respectively. Upon the optimisation of sulfur passivation parameters and subsequent encapsulation with atomic layer deposition Al₂O₃, the surface resistivity significantly increased from 4.3 kΩ.cm to 28.6 kΩ.cm, leading to a 19.1 times drop in dark current at room temperature for the GaSb p–i–n structure. This Letter provides a repeatable and stable passivation approach for improving the optoelectronic performance of GaSb-based devices.

The technical performance requirements from wireless communication networks are continuously rising. One method to satisfy the demands is increasing the carrier frequency to the millimetre wave or low terahertz band spectrum to utilise wider operation bandwidth. In order to facilitate the studies in this frequency range, the corresponding electromagnetic (EM) wave properties, channel attributes and material characteristics need to be analytically formulated. In line with these, this Letter initially presents the theoretical expressions governing the EM wave transmission across a conducting medium. Then, by using the relative parameter quantities in the proposed attenuation constant () computation technique, the results of the measurements performed between 260 and 350 GHz for the clear window glass samples of different thicknesses are given. This Letter concludes with the evaluation of the outcomes.

This Letter proposed a maximum-a-posteriori-based, pilot-aided algorithm for iterative estimation of carrier frequency offset (CFO) and time-varying multiple-input multiple-output (MIMO) multipath channel responses in FD orthogonal frequency division multiplexing (OFDM) systems. Empirical results and related benchmark limits show that the proposed algorithm offers reliable system performance having (i) estimation's mean-squared errors close to related Bayesian Cramér Rao bounds, (ii) data-detection's bit-error rates close to the ideal values under the assumption of perfect channel estimation and synchronisation and (iii) high robustness against the time variation of wireless channels in a network with moving nodes.