Manual Feature Extraction Research Articles

Pruning- and Quantization-Based Compression Algorithm for Number of Mixed Signals Identification Network

Source number estimation plays an important role in successful blind signal separation. At present, the application of machine learning allows the processing of signals without the time-consuming and complex work of manual feature extraction. However, the convolutional neural network (CNN) for processing complex signals has some problems, such as incomplete feature extraction and high resource consumption. In this paper, a lightweight source number estimation network (LSNEN), which can achieve a robust estimation of the number of mixed complex signals at low SNR (signal-to-noise ratio), is studied. Compared with other estimation methods, which require manual feature extraction, our network can realize the extraction of the depth feature of the original signal data. The convolutional neural network realizes complex mapping of modulated signals through the cascade of multiple three-dimensional convolutional modules. By using a three-dimensional convolution module, the mapping of complex signal convolution is realized. In order to deploy the network in the mobile terminal with limited resources, we further propose a compression method for the network. Firstly, the sparse structure network is obtained by the weight pruning method to accelerate the speed of network reasoning. Then, the weights and activation values of the network are quantified at a fixed point with the method of parameter quantization. Finally, a lightweight network for source number estimation was obtained, which was compressed from 12.92 MB to 3.78 MB with a compression rate of 70.74%, while achieving an accuracy of 94.4%. Compared with other estimation methods, the lightweight source number estimation network method proposed in this paper has higher accuracy, less model space occupation, and can realize the deployment of the mobile terminal.

Improvement of Classification Accuracy of Four-Class Voluntary-Imagery fNIRS Signals using Convolutional Neural Networks

Multiclass functional Near-Infrared Spectroscopy (fNIRS) signal classification has become a convenient way for optical brain-computer interface. fNIRS signal classification with high accuracy is a challenging assignment while the signals are produced by means of voluntary and imagery movements of the same limb. Since the activation in time and space of voluntary and imagery movement show a similar pattern, the classification accuracy by the conventional shallow classifiers cannot reach an acceptable range. This paper proposes an accuracy improvement approach with the use of Convolutional Neural Networks (CNNs). In this work, voluntary and imagery hand movements (left hand and right hand) were performed by several participants. These four-class signals were acquired utilizing fNIRS devices. The signals were separated based on the tasks and filtered. With manual feature extraction, the signals were classified by support vector machine and linear discriminant analysis. The automatic feature extraction and classification mechanism of the CNN were applied to the fNIRS signals. From the results, it was found that CNN improves the classification accuracy to an acceptable range, which has not been achieved by any convolutional network.

A Data-Centric Approach for Portuguese Speech Recognition: Language Model And Its Implications

Recent advances in Automatic Speech Recognition have made it possible to achieve a quality never seen before in the literature, both for languages with abundant data, such as English, which has a large number of studies and for the Portuguese language, which has a more limited amount of resources and studies. The most recent advances address speech recognition problems with Transformers based models, which have the capability to perform the speech recognition task directly from the raw signal, without the need for manual feature extraction. Some studies have already shown that it is possible to further improve the quality of the transcription of these models using language models within the decoding stage, however, the real impact of such language models is still not clear, especially for the Brazilian Portuguese scenario. Also, it is known that the quality of the data used for training the models is of paramount importance, however, there are few works in the literature addressing this issue. This work explores the impact of language models applied to Portuguese speech recognition both in terms of data quality and computational performance, with a data-centric approach. We propose an approach to measure similarity between datasets and, thus, assist in decision-making during training. The approach indicates paths for the advancement of the state-of-the-art aiming at Portuguese speech recognition, showing that it is possible to reduce the size of the language model by 80% and still achieve error rates around 7.17% for the Common Voice dataset. The source code is available at https://github.com/joaoalvarenga/language-model-evaluation.

Disruption prediction based on fusion feature extractor on J-TEXT

Predicting disruptions across different tokamaks is necessary for next generation device. Future large-scale tokamaks can hardly tolerate disruptions at high performance discharge, which makes it difficult for current data-driven methods to obtain an acceptable result. A machine learning method capable of transferring a disruption prediction model trained on one tokamak to another is required to solve the problem. The key is a feature extractor which is able to extract common disruption precursor traces in tokamak diagnostic data, and can be easily transferred to other tokamaks. Based on the concerns above, this paper presents a deep feature extractor, namely, the fusion feature extractor (FFE), which is designed specifically for extracting disruption precursor features from common diagnostics on tokamaks. Furthermore, an FFE-based disruption predictor on J-TEXT is demonstrated. The feature extractor is aimed to extracting disruption-related precursors and is designed according to the precursors of disruption and their representations in common tokamak diagnostics. Strong inductive bias on tokamak diagnostics data is introduced. The paper presents the evolution of the neural network feature extractor and its comparison against general deep neural networks, as well as a physics-based feature extraction with a traditional machine learning method. Results demonstrate that the FFE may reach a similar effect with physics-guided manual feature extraction, and obtain a better result compared with other deep learning methods.

Exploring deep residual network based features for automatic schizophrenia detection from EEG

Schizophrenia is a severe mental illness which can cause lifelong disability. Most recent studies on the Electroencephalogram (EEG)-based diagnosis of schizophrenia rely on bespoke/hand-crafted feature extraction techniques. Traditional manual feature extraction methods are time-consuming, imprecise, and have a limited ability to balance accuracy and efficiency. Addressing this issue, this study introduces a deep residual network (deep ResNet) based feature extraction design that can automatically extract representative features from EEG signal data for identifying schizophrenia. This proposed method consists of three stages: signal pre-processing by average filtering method, extraction of hidden patterns of EEG signals by deep ResNet, and classification of schizophrenia by softmax layer. To assess the performance of the obtained deep features, ResNet softmax classifier and also several machine learning (ML) techniques are applied on the same feature set. The experimental results for a Kaggle schizophrenia EEG dataset show that the deep features with support vector machine classifier could achieve the highest performances (99.23% accuracy) compared to the ResNet classifier. Furthermore, the proposed model performs better than the existing approaches. The findings suggest that our proposed strategy has capability to discover important biomarkers for automatic diagnosis of schizophrenia from EEG, which will aid in the development of a computer assisted diagnostic system by specialists.

Automatic Detection of Basal Units beneath Antarctic Ice Sheet in Radargram Based on Deep Learning

Sea level rise, caused by accelerated melting of glaciers in Greenland and Antarctica in recent decades, has become a major concern in scientific, environmental, and political arenas. A comprehensive study of subglacial conditions and processes is particularly important for reliable analysis about their future evolution. Basal units – visibly distinct englacial structures near ice-bedrock interface, could provide substantial insight into subglacial processes, ice-sheet dynamic history and dramatically influence the flow of surrounding ice. In order to enable improved characterization of these features, we develop and apply an algorithm that allows for automatic detection of basal units, therefore, an algorithm based on deep learning is proposed. Compared with traditional method based on manual feature extraction, proposed algorithm can achieve completely automatic recognition of basal units in radargram. The network is mainly composed of ResNet and ASPP module, which can achieve high accuracy in very small dataset. We use radar data collected by Polar Geophysics Group (PGG) of The Earth Institute at Columbia University in Antarctica in 2008-2009 and 2009-2010 for experiments, results confirm the effectiveness and robustness of proposed algorithm. At the same time, a more rapid deep learning method is tried, which uses the lightweight network MobileNet V3 as backbone, obtaining a network structure that can save 81.8 percent of training time and 92.2 percent of processing time with high accuracy. It provides a possibility for rapid network training, application on mobile devices and real-time processing of radargram.

DC–DC​ circuit fault diagnosis based on GWO optimization of 1DCNN-GRU network hyperparameters

Aiming at the problem that traditional machine learning methods rely on manual feature extraction in DC–DC circuit soft faults, it is important to effectively obtain the fault characteristics of DC–DC circuit soft faults. In this study, combining the advantages of one-dimensional convolutional neural network (1DCNN) and gated logic unit (GRU), a deep learning model for fault identification of time series signals is proposed to realize soft fault diagnosis of DC–DC​ circuits. In this study, a 1DCNN-GRU network model is constructed, and 1DCNN can directly perform automatic feature extraction on the data, while the GRU makes up for the shortcomings of CNN in processing time series data, thereby ensuring the comprehensiveness of the extracted features. For the hyperparameter problem in the network model, the powerful optimization ability of the gray wolf optimization algorithm is used to automatically search for the best hyperparameters in the 1DCNN-GRU network, and then the optimized 1DCNN-GRU network model is used for comprehensive feature learning. In order to meet the needs of deep learning for large data samples, the overlapping sampling method is used to enrich the data sample set. Experimental results show that the proposed method achieves 99.62% accuracy in DC–DC circuit fault diagnosis, and still maintains good robustness in noisy environment.

Working Mode Recognition of Non-Specific Radar Based on ResNet-SVM Learning Framework.

Mode recognition is a basic task to interpret the behavior of multi-functional radar. The existing methods need to train complex and huge neural networks to improve the recognition ability, and it is difficult to deal with the mismatch between the training set and the test set. In this paper, a learning framework based on residual neural network (ResNet) and support vector machine (SVM) is designed, to solve the problem of mode recognition for non-specific radar, called multi-source joint recognition framework (MSJR). The key idea of the framework is to embed the prior knowledge of radar mode into the machine learning model, and combine the manual intervention and automatic extraction of features. The model can purposefully learn the feature representation of the signal on the working mode, which weakens the impact brought by the mismatch between training and test data. In order to solve the problem of difficult recognition under signal defect conditions, a two-stage cascade training method is designed, to give full play to the data representation ability of ResNet and the high-dimensional feature classification ability of SVM. Experiments show that the average recognition rate of the proposed model, with embedded radar knowledge, is improved by 33.7% compared with the purely data-driven model. Compared with other similar state-of-the-art reported models, such as AlexNet, VGGNet, LeNet, ResNet, and ConvNet, the recognition rate is increased by 12%. Under the condition of 0-35% leaky pulses in the independent test set, MSJR still has a recognition rate of more than 90%, which also proves its effectiveness and robustness in the recognition of unknown signals with similar semantic characteristics.

Vehicle make and model recognition using mixed sample data augmentation techniques

<span lang="EN-US">Vehicle identification based on make and model is an integral part of an intelligent transport system that helps traffic monitoring and crime control. Much research has been performed in this regard, but most of them used manual feature extraction or ensemble convolution neural networks that result in increased execution time during inference. This paper compared three deep learning models and utilized different augmentation techniques to achieve state-of-the-art performance without ensembling or fusing the models. Experimentations are made without any augmentation, with standard augmentation, and by mixed sample data augmentation techniques. Gradient accumulation and stochastic weighted averaging with mixed precision are used to have a large batch size that helped to reduce training time. The dataset comprised 48 vehicles’ models running on the road of Pakistan. The highest accuracy and F1 score of 97% and 95% using the FMix augmentation technique with EfficientNetV2-S architecture gave the confidence that the proposed solution can be implemented in production. </span>

GLADS: A global-local attention data selection model for multimodal multitask encrypted traffic classification of IoT

With the rapid development of the Internet of Things (IoT), numerous of IoT devices and different characteristics in IoT traffic patterns need traffic classification to enable many important applications. Deep-learning-based (DL-based) traffic methods have gained increasing attention due to their high accuracy and because manual feature extraction is not needed. Furthermore, seek a lightweight, multitask methods that supports a “performance-speed” trade-off. Thus, we proposed the 0.11 M global-local attention data selection (GLADS) model. The core of the GLADS model includes an “indicator” mechanism and a “local + global” framework. The “indicator” mechanism is a completely different method for handling multimodal input that allows the model to efficiently extract features from multimodal input with a single-modal-like approach. The “local + global” framework for the “performance-speed” trade-off includes a “local” part to obtain the features of each patch in the model input and a Global-Local Attention mechanism in the “global” part outputs the classification results under all possible lengths. Tests on the ISCX-VPN-2016, ISCX-Tor-2016, USTC-TFC-2016, and TON_IoT datasets show that GLADS achieves better performance than several state-of-the-art baselines, ranging from 2.42% to 7.76%. Furthermore, we also propose the “indicator,” which allows the model to simply cope with multimodal input. Based on global-local attention, we analyze the relation of the input section and model performance in detail.

A Flexible deep convolutional neural network coupled with progressive training framework for online capacity estimation of lithium-ion batteries

Machine learning-based methods have shown great application prospects on the capacity estimation of lithium-ion battery. However, most of extant approaches require complete charging and discharging profiles that rarely occurs in practical applications to extract health features matched by the cycling capacity. To address this limitation, this paper directly intercepted partial charging voltage segments from raw charging curves, evading complicated manual features extraction. A progressive training framework is proposed and integrated with the deep convolutional neural network to build the online capacity estimator. Three battery aging datasets involving 28 cells cyclic aging data are exploited for model training and validation. Thereinto, the source dataset and target dataset-1 are used to train the model parameters progressively, and the target dataset-2 is applied to validate model performance. Experimental results show that the deep convolutional neural network coupled with progressive training framework can significantly enhance the capacity estimation accuracy and reduce retraining time compared with the ordinary convolutional neural network. The comparison results of different machine learning algorithms demonstrate that the proposed method possesses high estimation precision with a maximum relative error of only 3.4% on the target dataset, which can be easily extended to different types of battery capacity estimation.

Deep Learning-Based Auto-Segmentation of Spinal Cord Internal Structure of Diffusion Tensor Imaging in Cervical Spondylotic Myelopathy.

Cervical spondylotic myelopathy (CSM) is a chronic disorder of the spinal cord. ROI-based features on diffusion tensor imaging (DTI) provide additional information about spinal cord status, which would benefit the diagnosis and prognosis of CSM. However, the manual extraction of the DTI-related features on multiple ROIs is time-consuming and laborious. In total, 1159 slices at cervical levels from 89 CSM patients were analyzed, and corresponding fractional anisotropy (FA) maps were calculated. Eight ROIs were drawn, covering both sides of lateral, dorsal, ventral, and gray matter. The UNet model was trained with the proposed heatmap distance loss for auto-segmentation. Mean Dice coefficients on the test dataset for dorsal, lateral, and ventral column and gray matter were 0.69, 0.67, 0.57, 0.54 on the left side and 0.68, 0.67, 0.59, 0.55 on the right side. The ROI-based mean FA value based on segmentation model strongly correlated with the value based on manual drawing. The percentages of the mean absolute error between the two values of multiple ROIs were 0.07, 0.07, 0.11, and 0.08 on the left side and 0.07, 0.1, 0.1, 0.11, and 0.07 on the right side. The proposed segmentation model has the potential to offer a more detailed spinal cord segmentation and would be beneficial for quantifying a more detailed status of the cervical spinal cord.

Accurate tumor segmentation and treatment outcome prediction with DeepTOP

BackgroundAccurate outcome prediction prior to treatment can facilitate trial design and clinical decision making to achieve better treatment outcome. MethodWe developed the DeepTOP tool with deep learning approach for region-of-interest segmentation and clinical outcome prediction using magnetic resonance imaging (MRI). DeepTOP was constructed with an automatic pipeline from tumor segmentation to outcome prediction. In DeepTOP, the segmentation model used U-Net with a codec structure, and the prediction model was built with a three-layer convolutional neural network. In addition, the weight distribution algorithm was developed and applied in the prediction model to optimize the performance of DeepTOP. ResultsA total of 1889 MRI slices from 99 patients in the phase III multicenter randomized clinical trial (NCT01211210) on neoadjuvant treatment for rectal cancer was used to train and validate DeepTOP. We systematically optimized and validated DeepTOP with multiple devised pipelines in the clinical trial, demonstrating a better performance than other competitive algorithms in accurate tumor segmentation (Dice coefficient: 0.79; IoU: 0.75; slice-specific sensitivity: 0.98) and predicting pathological complete response to chemo/radiotherapy (accuracy: 0.789; specificity: 0.725; and sensitivity: 0.812). DeepTOP is a deep learning tool that could avoid manual labeling and feature extraction and realize automatic tumor segmentation and treatment outcome prediction by using the original MRI images. ConclusionDeepTOP is open to provide a tractable framework for the development of other segmentation and predicting tools in clinical settings. DeepTOP-based tumor assessment can provide a reference for clinical decision making and facilitate imaging marker-driven trial design.

FASFLNet: feature adaptive selection and fusion lightweight network for RGB-D indoor scene parsing.

RGB-D indoor scene parsing is a challenging task in computer vision. Conventional scene-parsing approaches based on manual feature extraction have proved inadequate in this area because indoor scenes are both unordered and complex. This study proposes a feature adaptive selection, and fusion lightweight network (FASFLNet) for RGB-D indoor scene parsing that is both efficient and accurate. The proposed FASFLNet utilizes a lightweight classification network (MobileNetV2), constituting the backbone of the feature extraction. This lightweight backbone model guarantees that FASFLNet is not only highly efficient but also provides good performance in terms of feature extraction. The additional information provided by depth images (specifically, spatial information such as the shape and scale of objects) is used in FASFLNet as supplemental information for feature-level adaptive fusion between the RGB and depth streams. Furthermore, during decoding, the features of different layers are fused from top-bottom and integrated at different layers for final pixel-level classification, resulting in an effect similar to that of pyramid supervision. Experimental results obtained on the NYU V2 and SUN RGB-D datasets indicate that the proposed FASFLNet outperforms existing state-of-the-art models and is both highly efficient and accurate.

ConvLSTM-Att: An Attention-Based Composite Deep Neural Network for Tool Wear Prediction

In order to improve the accuracy of tool wear prediction, an attention-based composite neural network, referred to as the ConvLSTM-Att model (1DCNN-LSTM-Attention), is proposed. Firstly, local multidimensional feature vectors are extracted with the help of a one-dimensional convolutional neural network (1D-CNN), which avoids the loss of wear features caused by manual feature extraction. Then the temporal relationship learning between multidimensional feature vectors is performed by introducing a long short-term memory (LSTM) network to make up for the lack of long-short distance dependence of the captured sequence of the CNN network. Finally, an attention mechanism is applied to strengthen the ability to extract key information from tool-wearing temporal features. The proposed ConvLSTM-Att model is trained with the measured tool wear data and then performs as a tool wear predictor. The model is compared with several state-of-the-art models on the PHM tool wear data sets. It significantly outperforms the other models in terms of prediction accuracy, but with similar computational complexity.

Voice Pathology Detection Using a Two-Level Classifier Based on Combined CNN–RNN Architecture

The construction of an automatic voice pathology detection system employing machine learning algorithms to study voice abnormalities is crucial for the early detection of voice pathologies and identifying the specific type of pathology from which patients suffer. This paper’s primary objective is to construct a deep learning model for accurate speech pathology identification. Manual audio feature extraction was employed as a foundation for the categorization process. Incorporating an additional piece of information, i.e., voice gender, via a two-level classifier model was the most critical aspect of this work. The first level determines whether the audio input is a male or female voice, and the second level determines whether the agent is pathological or healthy. Similar to the bulk of earlier efforts, the current study analyzed the audio signal by focusing solely on a single vowel, such as /a/, and ignoring phrases and other vowels. The analysis was performed on the Saarbruecken Voice Database,. The two-level cascaded model attained an accuracy and F1 score of 88.84% and 87.39%, respectively, which was superior to earlier attempts on the same dataset and provides a steppingstone towards a more precise early diagnosis of voice complications.

T-type inverter fault diagnosis based on GASF and improved AlexNet

Aiming at the problems of high similarity of T-type inverter open-circuit fault features, cumbersome manual extraction of fault features, and the inability of one-dimensional convolutional neural network to fully play the role of feature extraction. In this paper, an end-to-end fault diagnosis method is proposed, which is based on Gramian Angular Summation Field and improved AlexNet network. The fault results can be diagnosed by collecting only the single-phase line voltage. Firstly, the collected one-dimensional timing signal is mapped into two-dimensional images through the Gram summation angle field algorithm. Then feature extraction is performed by the improved AlexNet to take advantage of its ability to extract image features. Finally, the fault diagnosis result is output by the Softmax layer. Simulation experiments show that the method proposed in this paper can automatically extract features that are helpful for fault identification from raw data. The fault diagnosis rate of this model is as high as 99.72%, and it can diagnose not only a single fault, but also multiple faults in different phases. Compared with other methods, the proposed method shows a better fault feature extraction effect and higher fault diagnosis accuracy.

Raman spectroscopy and convolutional neural networks for monitoring biochemical radiation response in breast tumour xenografts

Tumour cells exhibit altered metabolic pathways that lead to radiation resistance and disease progression. Raman spectroscopy (RS) is a label-free optical modality that can monitor post-irradiation biomolecular signatures in tumour cells and tissues. Convolutional Neural Networks (CNN) perform automated feature extraction directly from data, with classification accuracy exceeding that of traditional machine learning, in cases where data is abundant and feature extraction is challenging. We are interested in developing a CNN-based predictive model to characterize clinical tumour response to radiation therapy based on their degree of radiosensitivity or radioresistance. In this work, a CNN architecture is built for identifying post-irradiation spectral changes in Raman spectra of tumour tissue. The model was trained to classify irradiated versus non-irradiated tissue using Raman spectra of breast tumour xenografts. The CNN effectively classified the tissue spectra, with accuracies exceeding 92.1% for data collected 3 days post-irradiation, and 85.0% at day 1 post-irradiation. Furthermore, the CNN was evaluated using a leave-one-out- (mouse, section or Raman map) validation approach to investigate its generalization to new test subjects. The CNN retained good predictive accuracy (average accuracies 83.7%, 91.4%, and 92.7%, respectively) when little to no information for a specific subject was given during training. Finally, the classification performance of the CNN was compared to that of a previously developed model based on group and basis restricted non-negative matrix factorization and random forest (GBR-NMF-RF) classification. We found that CNN yielded higher classification accuracy, sensitivity, and specificity in mice assessed 3 days post-irradiation, as compared with the GBR-NMF-RF approach. Overall, the CNN can detect biochemical spectral changes in tumour tissue at an early time point following irradiation, without the need for previous manual feature extraction. This study lays the foundation for developing a predictive framework for patient radiation response monitoring.

A refinement network embedded with attention mechanism for computer vision based post-earthquake inspections of railway viaduct

Post-earthquake inspection of structures based on computer vision is developing rapidly due to the advantages of high efficiency and without manual feature extraction. However, it is still necessary to investigate how to accurately recognize structural components and damage from the perspective of pixels. Fortunately, refinement network which named RefineNet has been developed for semantic segmentation of images, which helps to combine low-level features and high-level semantics to generate high resolution segmented images for efficient end-to-end learning. Therefore, RefineNet is used in this study as a network architecture for semantic segmentation tasks of recognizing railway viaducts components and damages. Moreover, it is proposed to embed the convolutional block attention mechanism in the down-sampling process of the RefineNet to extract image features, which helps the network to assign different weights to image regions of different importance and effectively improve the extraction effect of intermediate features. With the provided large-scale synthetic railway viaduct image dataset, which named Tokaido Dataset, the proposed RefineNet with Attention Mechanism (RefineNet-AM) is used for structural condition assessment of railway viaduct, including semantic segmentation tasks of components and damages of railway viaduct. Based on the test dataset, it is shown that proposed RefineNet-AM can inspect the structural components and damage of railway viaduct with satisfactory accuracy.

AgriDet: Plant Leaf Disease severity classification using agriculture detection framework

In the field of modern agriculture, plant disease detection plays a vital role in improving crop productivity. To increase the yield on a large scale, it is necessary to predict the onset of the disease and give advice to farmers. Previous methods for detecting plant diseases rely on manual feature extraction, which is more expensive. Therefore, image-based techniques are gaining interest in the research area of plant disease detection. However, existing methods have several problems due to the improper nature of the captured image, including improper background conditions that lead to occlusion, illumination, orientation, and size. Also, cost complexity, misclassifications, and overfitting problems occur in several real-time applications. To solve these issues, we proposed an Agriculture Detection (AgriDet) framework that incorporates conventional Inception-Visual Geometry Group Network (INC-VGGN) and Kohonen-based deep learning networks to detect plant diseases and classify the severity level of diseased plants. In this framework, image pre-processing is done to remove all the constraints in the captured image. Then, the occlusion problem is tackled by the proposed multi-variate grabcut algorithm for effective segmentation. Furthermore, the framework performs accurate disease detection and classification by utilizing an improved base network, namely a pre-trained conventionally based INC-VGGN model. Here, the pre-trained INC-VGGN model is a deep convolutional neural network for prediction of plant diseases that was previously trained for the distinctive dataset. The pre-trained weights and the features learned in this base network are transferred into the newly developed neural network to perform the specific task of plant disease detection for our dataset. In order to overcome the overfitting problem, a dropout layer is introduced, and the deep learning of features is performed using the Kohonen learning layer. After percentage computation, the improved base network classifies the severity classes in the training sets. Finally, the performance of the framework is computed for different performance metrics and achieves better accuracy than previous models. Also, the performance of the statistical analysis is validated to prove the results in terms of accuracy, specificity, and sensitivity.