Feature Extraction Methodology Research Articles

An empirical study of handcrafted and dense feature extraction techniques for lung and colon cancer classification from histopathological images

According to a 2020 WHO report, cancer is one of the main causes of deaths worldwide. Among these deaths, lung and colon cancer collectively responsible for nearly 2.735 million deaths.So, detection and classification of lung and colon cancer is one of the utmost priority research areas in the field of biomedical health informatics. In this article, comparative analysis of two feature extraction methodologies has been presented for lung and colon cancer classification. In one approach, six handcrafted features extraction techniques based on colour, texture, shape and structure are presented. Gradient Boosting (GB), SVM-RBF, Multilayer Perceptron (MLP) and Random Forest (RF) classifiers with handcrafted features are trained and tested for lung and colon cancer classification. In another approach, using the notion of transfer learning, seven deep learning frameworks for deep feature extraction from lung and colon cancer histopathological images are presented. The extracted deep features (as input attributes) are applied into conventional GB, SVM-RBF, MLP and RF classifiers for lung and colon cancer classification. However, in contrast to handcrafted features a significant improvement in classifiers performance is observed with features extracted by deep CNN networks. It has been found that the proposed technique obtained excellent results in terms of accuracy, precision, recall, F1 score and ROC-AUC. The RF classifier with DenseNet-121 extracted deep features can identify the lung and colon cancer tissue with an accuracy and recall of 98.60%, precision of 98.63%, F1 score of 0.985 and ROC-AUC of 01.

Intelligent feature extraction, data fusion and detection of concrete bridge cracks: current development and challenges

As a common appearance defect of concrete bridges, cracks are important indices for bridge structure health assessment. Although there has been much research on crack identification, research on the evolution mechanism of bridge cracks is still far from practical applications. In this paper, the state-of-the-art research on intelligent theories and methodologies for intelligent feature extraction, data fusion and crack detection based on data-driven approaches is comprehensively reviewed. The research is discussed from three aspects: the feature extraction level of the multimodal parameters of bridge cracks, the description level and the diagnosis level of the bridge crack damage states. We focus on previous research concerning the quantitative characterization problems of multimodal parameters of bridge cracks and their implementation in crack identification, while highlighting some of their major drawbacks. In addition, the current challenges and potential future research directions are discussed.

GET A LIST OF FEATURE EXTRACTORS BASED ON FEATURE IMPORTANCE TECHNIQUES

It is well-known, we live in the digital industrial era and we have a lot of accumulated volumes of structured and unstructured data. The accumulated data, actually, can help us make our solutions more efficient in different areas of human activities. That is why we are trying to recognize which kind of data can resolve which kind of problem. For this purpose, we have a lot of useful techniques for assessment of the data feature importance. We are more confident in feature importance techniques when we work with numerical values of structured data because the detected objects are described by particular values. Alas, when extracting the features from unstructured data, like from images, it is still a question of how feature importance techniques are useful because it is either a question of how we are going to describe a detected object or which set of features might be an optimal one. Of course, it is a good case when we can extract a lot of different features of a detected object but how it is enough if we want to recognize it from a list of other possible objects? This problem we are going to resolve this by an approach where we describe the nature of the detected object via its unique features either than feature importance techniques. In the article, we consider the state of art feature importance methods, feature extraction algorithms and methodology of how to create the optimal set of feature extractors. Keywords: Computer vision, object detection, feature importance, machine learning, water pollution, liquid contamination.

Defect Detection in Printed Circuit Boards with Pre-Trained Feature Extraction Methodology with Convolution Neural Networks

Printed Circuit Boards (PCBs) are very important for proper functioning of any electronic device. PCBs are installed in almost all the electronic device and their functionality is dependent on the perfection of PCBs. If PCBs do not function properly then the whole electric machine might fail. So, keeping this in mind researchers are working in this field to develop error free PCBs. Initially these PCBs were examined by the human beings manually, but the human error did not give good results as sometime defected PCBs were categorized as non-defective. So, researchers and experts transformed this manual traditional examination to automated systems. Further to this research image processing and computer vision came into actions where the computer vision experts applied image processing techniques to extract the defects. But, this also did not yield good results. So, to further explore this area Machine Learning and Artificial Intelligence Techniques were applied. In this study we have applied Deep Neural Networks to detect the defects in the PCBS. Pretrained VGG16 and Inception networks were applied to extract the relevant features. DeepPCB dataset was used in this study, it has 1500 pairs of both defected and non-defected images. Image pre-processing and data augmentation techniques were applied to increase the training set. Convolution neural networks were applied to classify the test data. The results were compared with state-of-the art technique and it proved that the proposed methodology outperformed it. Performance evaluation metrics were applied to evaluate the proposed methodology. Precision 94.11%, Recall 89.23%, F-Measure 91.91%, and Accuracy 92.67%.

Cell population-based framework of genetic epidemiology in the single-cell omics era.

Genetic epidemiology is a rapidly advancing field due to the recent availability of large amounts of omics data. In recent years, it has become possible to obtain omics information at the single-cell level, so genetic epidemiological models need to be updated to integrate with single-cell expression data. In this perspective paper, we propose a cell population-based framework for genetic epidemiology in the single-cell era. In this framework, genetic diversity influences phenotypic diversity through the diversity of cell population profiles, which are defined as high-dimensional probability distributions of the state spaces of biomolecules of each omics layer. We discuss how biomolecular experimental measurement data can capture the different properties of this distribution. In particular, single-cell data constitute a sample from this population distribution where only some coordinate values are observable. From a data analysis standpoint, we introduce methodology for feature extraction from cell population profiles. Finally, we discuss how this framework can be applied not only to genetic epidemiology but also to systems biology.

Analysis of Performance of Two Wavelet Families Using GLCM Feature Extraction for Mammogram Classification of Breast Cancer

Background: Mammogram images are low dose x-ray images which detect the breast cancer before the women can actually experience it. Objective: To determine the accurate methodology for feature extraction using different wavelet families and different classification algorithms. Methods: Two wavelet families are used namely Daubechies (db8) and Biorthogonal (bior3.7). The Gray-Level Co-occurrence Matrix is used for extracting 9 features at each sub-band. 27 features are extracted at three sub-bands of Discrete Wavelet Transform. The features are extracted at three levels of decomposition and after that the classification algorithm named as Naive Bayes, Multilayer Perceptron, Fuzzy-NN and Genetic Programming are applied to extracted features. The feature selection algorithms are applied named as Wavelet and Principle Component Analysis for selecting the features and then classification accuracy is determined and compared between these. Results: Mammographic Image Analysis Society, database including 322 mammogram images from 161 patients is used. The classification algorithm without feature selection named as Fuzzy-NN gives better results at the third level of decomposition having classification accuracy for db8 wavelet family up to 99.68% and for bior3.7 wavelet family up to 99.98%. Wavelet with Multilayer Perceptron using feature selection algorithm gives the classification accuracy for db8 wavelet family up to 96.27% and for bior3.7 up to 93.47%. Conclusion: Fuzzy-NN algorithm gives highest accuracy of 99.98% for bior3.7 wavelet family. It indicates that with feature selection and without feature selection, the wavelet families differ as db8 is better consideration for with feature selection and bior3.7 wavelet family for without feature selection.

A CNN-LSTM network with multi-level feature extraction-based approach for automated detection of coronavirus from CT scan and X-ray images

Auto-detection of diseases has become a prime issue in medical sciences as population density is fast growing. An intelligent framework for disease detection helps physicians identify illnesses, give reliable and consistent results, and reduce death rates. Coronavirus (Covid-19) has recently been one of the most severe and acute diseases in the world. An automatic detection framework should therefore be introduced as the fastest diagnostic alternative to avoid Covid-19 spread. In this paper, an automatic Covid-19 identification in the CT scan and chest X-ray is obtained with the help of a combined deep learning and multi-level feature extraction methodology. In this method, the multi-level feature extraction approach comprises GIST, Scale Invariant Feature Transform (SIFT), and Convolutional Neural Network (CNN) extract features from CT scans and chest X-rays. The objective of multi-level feature extraction is to reduce the training complexity of CNN network, which significantly assists in accurate and robust Covid-19 identification. Finally, Long Short-Term Memory (LSTM) along the CNN network is used to detect the extracted Covid-19 features. The Kaggle SARS-CoV-2 CT scan dataset and the Italian SIRM Covid-19 CT scan and chest X-ray dataset were employed for testing purposes. Experimental outcomes show that proposed approach obtained 98.94% accuracy with the SARS-CoV-2 CT scan dataset and 83.03% accuracy with the SIRM Covid-19 CT scan and chest X-ray dataset. The proposed approach helps radiologists and practitioners to detect and treat Covid-19 cases effectively over the pandemic.

Assessment of plant species distribution and diversity along a climatic gradient from Mediterranean woodlands to semi-arid shrublands

ABSTRACT Climate and land-use change profoundly affect plant species distribution (SD) and composition, and the impact of these processes is expected to increase in the coming years. As a proxy of global changes, knowledge of SD and diversity along climatic gradients is essential to determine the efforts needed for species conservation. Plant spectral diversity is an emerging approach used as a proxy for species diversity based on remote sensing. Thus, the research aim was to develop a comprehensive methodology based on spectral diversity for SD and richness mapping and to study their relations with environmental and human-derived factors, demonstrated along Mediterranean to semi-arid climatic gradient. The study addresses two main knowledge gaps regarding spectral diversity: (1) improving the accuracy of woody species classification by features extraction and selection, and by using texture analysis in an ecosystem characterized by high spatial variability and relatively small-sized and sparse woody vegetation; and (2) developing a better estimate of the local species ‎richness and their response to environmental and human-derived factors (i.e. climate, topography, substrate, and land cover factors) across a transition zone between Mediterranean woodlands and semi-arid dwarf shrublands. A hyperspectral image was acquired for a 43-km strip along the study area using an airborne flight of AISA-FENIX (380–2500 nm, 420 bands) at the end of the 2017 rainy season. The dominant species were surveyed, with a total number of 247 trees and shrubs, to train a machine learning support vector machine (SVM) classification for species distribution mapping, which yielded an overall accuracy of 86.1%. A feature extraction and selection methodology was developed, combining principal component analysis and neighborhood component analysis techniques, facilitating the identification of 33 spectral diagnostic bands out of 330 spectral bands. The classification accuracy was decreased by about 2% to 84.2% using only 33 spectral bands. The classification accuracy improved by about 7.1% for the seven large crown species (93.3%) by adding texture information. Later, the local species richness was calculated by utilizing the alpha diversity index (i.e. the Shannon Index) for 30-m grid cells and was tested in response to environmental (i.e. climate, substrate, and topography) and human-derived factors (i.e. land cover). The highest sensitivity to alpha diversity factors was mean annual precipitation, slope, and land surface temperature. The alpha diversity showed higher richness in the natural Mediterranean shrubland and the guarrigue located in the northern part of the climate gradient. We suggest that the approach presented here significantly improves the estimation of woody species distribution and diversity in areas characterized by high spatial heterogeneity along steep climatic gradients.

A hybrid of deep and textural features to differentiate glomerulosclerosis and minimal change disease from glomerulus biopsy images

The minimal change disease (MCD) and glomerulosclerosis (GS) are two common kidney diseases. Unless adequately treated, these diseases leads to chronic kidney diseases. Accurate differentiation of these two diseases is of paramount importance as their methods of treatment and prognoses are different. Thus, this article propose a method capable of differentiating MCD from GS in glomerulus biopsies images based on a new hybrid deep and texture feature space. We conducted an extensive study to determine the best set of features for image representation. Our feature extraction methodology, which includes Haraliks and geostatistics texture descriptors and pre-trained CNNs, resulted in 13,476 characteristics. We then used mutual information to order the elements by importance and select the best set for differentiating MCD from GS using the random forest classifier. The proposed method achieved an accuracy of 90.3% and a Kappa index of 80.5%. Representation of glomerulus biopsy images with a hybrid of deep and textural features facilitates the accurate differentiation of GS and MCD.

Design of Accurate Classification of COVID-19 Disease in X-Ray Images Using Deep Learning Approach

COVID-19 appears to be having a devastating influence on world health and well-being. Moreover, the COVID-19 confirmed cases have recently increased to over 10 million worldwide. As the number of verified cases increase, it is more important to monitor and classify healthy and infected people in a timely and accurate manner. Many existing detection methods have failed to detect viral patterns. Henceforth, by using COVID-19 thoracic x-rays and the histogram-oriented gradients (HOG) feature extraction methodology; this research work has created an accurate classification method for performing a reliable detection of COVID-19 viral patterns. Further, the proposed classification model provides good results by leveraging accurate classification of COVID-19 disease based on the medical images. Besides, the performance of our proposed CNN classification method for medical imaging has been assessed based on different edge-based neural networks. Whenever there is an increasing number of a class in the training network, the accuracy of tertiary classification with CNN will be decreasing. Moreover, the analysis of 10 fold cross-validation with confusion metrics can also take place in our research work to detect various diseases caused due to lung infection such as Pneumonia corona virus-positive or negative. The proposed CNN model has been trained and tested with a public X-ray dataset, which is recently published for tertiary and normal classification purposes. For the instance transfer learning, the proposed model has achieved 85% accuracy of tertiary classification that includes normal, COVID-19 positive and Pneumonia. The proposed algorithm obtains good classification accuracy during binary classification procedure integrated with the transfer learning method.

Hardware Acceleration of EEG-Based Emotion Classification Systems: A Comprehensive Survey.

Recent years have witnessed a growing interest in EEG-based wearable classifiers of emotions, which could enable the real-time monitoring of patients suffering from neurological disorders such as Amyotrophic Lateral Sclerosis (ALS), Autism Spectrum Disorder (ASD), or Alzheimer's. The hope is that such wearable emotion classifiers would facilitate the patients' social integration and lead to improved healthcare outcomes for them and their loved ones. Yet in spite of their direct relevance to neuro-medicine, the hardware platforms for emotion classification have yet to fill up some important gaps in their various approaches to emotion classification in a healthcare context. In this paper, we present the first hardware-focused critical review of EEG-based wearable classifiers of emotions and survey their implementation perspectives, their algorithmic foundations, and their feature extraction methodologies. We further provide a neuroscience-based analysis of current hardware accelerators of emotion classifiers and use it to map out several research opportunities, including multi-modal hardware platforms, accelerators with tightly-coupled cores operating robustly in the near/supra-threshold region, and pre-processing libraries for universal EEG-based datasets.

Understanding gender differences in professional European football through machine learning interpretability and match actions data

After the great success of the Women’s World Cup in 2019, several platforms have started identifying the reasons for gender inequality in European football. Even though these inequalities emerge from a variety of key aspects in the modern sport, we focused on the game and evaluated the main differential features of European male and female football players in match actions data under the assumption of finding significant differences and established patterns between genders. A methodology for unbiased feature extraction and objective analysis is presented based on data integration and machine learning explainability algorithms. Female (n_0 = 1511) and male (n_1 = 2703) data points were collected from event data and categorized by game period and player position. We set up a supervised classification pipeline to predict the gender of each player by looking at their actions in the game. The comparison methodology did not include any qualitative enrichment or subjective analysis to prevent biased data enhancement or gender-related processing. The pipeline included three representative binary classification models; A logic-based Decision Trees, a probabilistic Logistic Regression and a multilevel perceptron Neural Network. Each model tried to draw the differences between male and female data points, and we extracted the results using machine learning explainability methods to understand the underlying mechanics of the models implemented. The study was able to determine pivotal factors that differentiate each gender performance as well as disseminate unique patterns by gender involving more than one indicator. Data enhancement and critical variables analysis are essential next steps to support this framework and serve as a baseline for further studies and training developments.

Ensemble Visual Content Based Search and Retrieval for Natural Scene Images

Background: Content Based Image Retrieval (CBIR) is one of the fields for information retrieval where similar images are retrieved from database based on the various image descriptive parameters. The image descriptor vector is used by machine learning based systems to store, learn and template matching. These feature descriptor vectors locally or globally demonstrate the visual content present in an image using texture, colour, shape, and other information. Objective: The main vision of this paper is to categorize and evaluate those algorithms, which were proposed in the interval of last 10 years. In addition, experiment is performed using a hybrid content descriptors methodology that helps to gain the significant results as compared with state-of-art algorithms. Methods: In past, several algorithms were proposed to fetch the variety of contents from an image based on which the image is retrieved from database. But, the precision and recall for the gained results using single content descriptor is not significant. The proposed system architecture uses the hybrid ensemble feature set for image matching. The hybrid parameters include globally and locally defined feature extraction methodologies. Results: The hybrid methodology decreases the error rate and improves the precision and recall for large natural scene images dataset having more than 20 classes. The overall combination will provide almost 97% accurate results which is better than the existing literature results. Conclusion: In conclusion the experimentation result suggests that the use of the local feature extraction mechanism is better when compared with the global feature extraction methodology.

A fast and efficient feature extraction methodology for structural damage localization based on raw acceleration measurements

Many damage detection strategies have been developed within the field of structural health monitoring, showing promising results in real-world applications. Most of them rely on the use of modal parameters to locate damage in structures. However, they are highly dependent on the process of extracting such characteristics, which contains uncertainties and needs to be adapted specifically for each structure. Recently, a special focus is being given to techniques based only on the use of raw structural acceleration measurements due to their relatively lower computational complexity manipulation. Thus, this work proposes an efficient feature extraction methodology for structural damage localization based on raw vibration signals. Furthermore, its ability to indicate damage quantification is also investigated. The proposed approach consists in extracting sensitive features from time, frequency, and quefrency domains. Then, percentile intervals are defined for each feature regarding the structures' healthy state. Finally, a damage index is estimated based on the number of outlier features of the structures' damaged state. In this sense, the sensor with the highest number of outlier features indicates the damage location. To assess this methodology, four different applications are studied: a numerical two-dimensional frame, a simply supported beam tested in laboratory, a three-dimensional frame also tested in laboratory, and the Tianjin Yonghe Bridge in China. Results show that the proposed approach is not only able to correctly indicate damage locations but also to give insights about their magnitudes.

Wavelet Transform Based Fault Identification and Reconfiguration for a Reduced Switch Multilevel Inverter Fed Induction Motor Drive

The multilevel inverter-based drive system is greatly affected by several faults occurring on switching elements. A faulty switch in the inverter can potentially lead to more losses, extensive downtime and reduced reliability. In this paper, a novel fault identification and reconfiguration process is proposed by using discrete wavelet transform and auxiliary switching cells. Here, the discrete wavelet transform exploits a multiresolution analysis with a feature extraction methodology for fault identification and subsequently for reconfiguration. For increasing the reliability, auxiliary switching cells are integrated to replace faulty cells in a proposed reduced-switch 5-level multilevel inverter topology. The novel reconfiguration scheme compensates open circuit and short circuit faults. The complexity of the proposed system is lower relative to existing methods. This proposed technique effectively identifies and classifies faults using the multiresolution analysis. Furthermore, the measured current and voltage values during fault reconfiguration are close to those under healthy conditions. The performance is verified using the MATLAB/Simulink platform and a hardware model.

Road Feature Extraction from LANDSAT-8 and ResourceSat-2 Images

This paper presents a methodology of road feature extraction from the different resolutions of Remote Sensing images of Landsat-8 Operational Lander Image (OLI) and ResourceSat-2 of Linear Imaging Self Sensor-3 (LISS-3) and LISS-4 sensors with the spatial resolutions of 15 m, 24 m, and 5 m. In the methodology of road extraction, an index is proposed based on the spectral profile of Roads, also involving Morphological transform (Top-Hat or Bot-Hat) and Markov Random Fields (MRF). In the proposed index, Short Wave Infrared (SWIR) band has a significant role in the detection of roads from sensors, and it is named Normalized Difference Road Index (NDRI). To enhancement of features from the index, Bot-Hat transforms used. To segment the road features from this image, MRF used. The methodology is performed on the OLI, LISS-3 and LISS-4 images, and presented with results.

DEELIG: A Deep Learning Approach to Predict Protein-Ligand Binding Affinity

Protein-ligand binding prediction has extensive biological significance. Binding affinity helps in understanding the degree of protein-ligand interactions and is a useful measure in drug design. Protein-ligand docking using virtual screening and molecular dynamic simulations are required to predict the binding affinity of a ligand to its cognate receptor. Performing such analyses to cover the entire chemical space of small molecules requires intense computational power. Recent developments using deep learning have enabled us to make sense of massive amounts of complex data sets where the ability of the model to “learn” intrinsic patterns in a complex plane of data is the strength of the approach. Here, we have incorporated convolutional neural networks to find spatial relationships among data to help us predict affinity of binding of proteins in whole superfamilies toward a diverse set of ligands without the need of a docked pose or complex as user input. The models were trained and validated using a stringent methodology for feature extraction. Our model performs better in comparison to some existing methods used widely and is suitable for predictions on high-resolution protein crystal (⩽2.5 Å) and nonpeptide ligand as individual inputs. Our approach to network construction and training on protein-ligand data set prepared in-house has yielded significant insights. We have also tested DEELIG on few COVID-19 main protease-inhibitor complexes relevant to the current public health scenario. DEELIG-based predictions can be incorporated in existing databases including RSCB PDB, PDBMoad, and PDBbind in filling missing binding affinity data for protein-ligand complexes.

Comparative Analysis of Feature Extraction and Pixel-based Classification of High-Resolution Satellite Images Using Geospatial Techniques

Two common techniques for classifying satellite imagery are pixel-based and Feature extraction image analysis methods. Typically, for agreements reached imaging, pixel-based analysis is used, whereas high-resolution imagery is suitable for Feature extraction analysis. However, In the classification of moderate images, image segmentation's ability depending on criteria such as shape, color, texture, and spatial features in Feature extraction image analysis implies it can perform better than pixel-based analysis. A comparative study of the two methods was performed using Sentinel-2 imagery from 18 May 2020 to categorize LU/LC in the City of Baghdad, Iraq. After calculating LU/LC for Baghdad images' capital, a supervised classification was performed using the two methods. The images used have been the support vector machines (SVM) and the maximum likelihood classification (MLC) for pixel-based method and Feature extraction method, which is available in ENVI and ArcGIS software packages, respectively. Land cover and land use classes included five Groups (vegetation area, asphalt road, soil area, water body, and built-up) was found that the Feature extraction methodology produced higher overall accuracy and Kappa index in the city of Baghdad image. The highest achieved accuracy for the Feature extraction technique was overall accuracy 95% with Kappa index 0.94 of SVM and overall accuracy of 92% with Kappa index 0.90 of MLC. In comparison, the highest accuracy for the pixel-based was overall accuracy 88% with Kappa index 0.84 of SVM and overall accuracy 86% with Kappa index 0.82 of MLC.

Drilling process monitoring: a framework for data gathering and feature extraction techniques

Today’s industrial transformation is taking advantage of the benefits of information and communication technologies (ICT) to evolve into a more decision-making environment in manufacturing. Efficiency, agility, innovation, quality and cost savings are the goals of this transformation in one of the most employed manufacturing processes as is the case of machining. Drilling processes are among the last operations in the different manufacturing stages of machined parts, where an undetected problem can lead to the production of a defective part. Data analysis of sensor signals gathered during drilling processes provides information related to the cutting process that can anticipate non-desired phenomena. This work illustrates the experimental setup for sensorial data acquisition in drilling processes, signal processing techniques and feature extraction methodologies for faster and more robust decision-making paradigms.

Palmvein Recognition Using Block-Based WLD Histogram of Gabor Feature Maps and Deep Neural Network With Bayesian Optimization

The demand on devices and systems empowered by biometric verification and identification mechanisms has been increasing in recent years as they have become a significant part of our lives. Palm vein biometric is an emerging technology that has drawn considerable attention from researchers and scientists over the last decade. In the present work, a novel feature extraction methodology named GPWLD combing the Gabor features with positional Weber's local descriptor (PWLD) features is proposed. WLD is a highly representative micro-pattern descriptor that performs well against noise and illumination changes in images. However, it lacks the ability to capture the vein pattern features at different orientations. Moreover, its descriptor packs the local information content into a single histogram that does not take the spatial locality of micro-patterns into consideration. To solve these two issues, firstly, the palm vein image is passed through Gabor filter with different orientations to capture the salient rotational features found in the output feature maps. Secondly, the spatiality is achieved by uniformly dividing each feature map into several blocks. Next, Weber's law feature descriptor (WLD) histogram is computed for each block in every feature map. Finally, these histograms are concatenated to compose the final feature vector. Due to the high dimensionality of the final feature vector, Principal component analysis (PCA) algorithm is utilized for feature size reduction. In the classification stage, a deep neural network (DNN) comprising an optimized stacked autoencoder (SAE) with Bayesian optimization and a softmax layer is used. Optimization of the SAE is carried out by using Bayesian optimization to find the optimal SAE structure and the options of the training algorithm. The Experimental results verify the discriminative power of the extracted features and the accuracy of the proposed DNN. For both Identification and verification experiments, the proposed method attains higher identification rate and lower EER than state-of-the-art methods.