Recognition Model Research Articles

Seeking optimal and explainable deep learning models for inertial-based posture recognition

Deep Learning (DL) models, widely used in several domains, are often applied for posture recognition. This work researches five DL architectures for posture recognition: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Transformer, hybrid CNN-LSTM, and hybrid CNN-Transformer. Agriculture and construction working postures were addressed as use cases, by acquiring an inertial dataset during the simulation of their typical tasks in circuits. Since model performance greatly depends on the choice of the hyperparameters, a grid search was conducted to find the optimal hyperparameters. An extensive analysis of the hyperparameter combinations’ effects is presented, identifying some general tendencies. Moreover, to unveil the black-box DL models, we applied the Gradient-weighted Class Activation Mapping (Grad-CAM) explainability method on CNN’s outputs to better understand the model’s decision-making, in terms of the most important sensors and time steps for each window output. Innovative hybrid architectures combining CNN and LSTM or Transformer encoder were implemented, by using the convolution feature maps as LSTM’s or Transformer’s inputs and fusing both subnetworks’ outputs with weights learned during the training. All architectures successfully recognized the eight posture classes, with the best model of each architecture exceeding 91.5% F1-score in the test. A top F1-score of 94.33%, with an inference time of just 0.29 ms (in a regular laptop), was achieved by a hybrid CNN-Transformer.

Deep learning utilization in Sundanese script recognition for cultural preservation

This study addresses the challenge of preserving the Sundanese script, a traditional writing system of the Sundanese community in Indonesia, which is at risk of being forgotten due to technological advancements. To tackle this problem, we propose a deep learning approach using the YOLOv8 model for the automatic recognition of Sundanese characters. Our methodology includes creating a comprehensive dataset, applying augmentation techniques, and annotating the characters. The trained model achieved a precision of 95% after 150 epochs, demonstrating its effectiveness in recognizing Sundanese characters. While some variability in accuracy was observed for certain characters and real-time applications, the results indicate the feasibility and promise of using deep learning for Sundanese script recognition. This research highlights the potential of technological solutions to digitize and preserve the Sundanese script, ensuring its continued legacy and accessibility for future generations. Thus, we contribute to cultural preservation by providing a method to safeguard the Sundanese script against obsolescence.

Efficient deep learning models for Telugu handwritten text recognition

<p>Optical character recognition (OCR) technology is indispensable for converting and analyzing text from various sources into a format that is editable and searchable. Telugu handwriting presents notable challenges due to the resemblance of characters, the extensive character set, and the need to segment overlapping characters. To segment the overlapping characters, we assess the width of small characters within a word and segment the overlapping characters accordingly. This method is well suited for the segmentation of overlapping compound characters. To address the recognition of similar characters with less training periods we have used ResNet 18 and SqueezeNet models which have achieved character recognition rates of 95% and 94% respectively.</p>

Automatic kidney disease prediction using deep learning techniques

The kidneys play an energetic role in eliminating excess products and fluids from the body, by a complex mechanism which is crucial for upholding a stable balance of body chemicals. Chronic kidney disease (CKD) is considered by an unhurried weakening in renal function that may eventually result in kidney injury or failure. The difficulty of diagnosing the illness rises as it worsens. However, using data from normal medical visits to evaluate the various phases of CKD could help with early detection and prompt care. Researchers suggest a classification strategy for CKD along with optimization strategies used in the learning process. The incorporation of artificial intelligence offers promise because it may often astonish with its skills and enable seemingly difficult undertakings. Modern machine learning techniques have been developed to detect renal illness in light of this. In the current study, a new deep learning model for CKD initial recognition and prediction is introduced. The main objective of the project is to build a strong deep neural network (DNN) and estimate its result outcomes in comparison to other leading-edge machine learning techniques. The outcomes demonstrate that the proposed strategy outperforms current approaches and has promise as a useful tool for CKD detection.

ALDANER: Active Learning based Data Augmentation for Named Entity Recognition

Training Named Entity Recognition (NER) models typically necessitates the use of extensively annotated datasets. This requirement presents a significant challenge due to the labor-intensive and costly nature of manual annotation, especially in specialized domains such as medicine and finance. To address data scarcity, two strategies have emerged as effective: (1) Active Learning (AL), which autonomously identifies samples that would most enhance model performance if annotated, and (2) data augmentation, which automatically generates new samples. However, while AL reduces human effort, it does not eliminate it entirely, and data augmentation often leads to incomplete and noisy annotations, presenting new hurdles in NER model training. In this study, we integrate AL principles into a data augmentation framework, named Active Learning-based Data Augmentation for NER (ALDANER), to prioritize the selection of informative samples from an augmented pool and mitigate the impact of noisy annotations. Our experiments across various benchmark datasets and few-shot scenarios demonstrate that our approach surpasses several data augmentation baselines, offering insights into promising avenues for future research.

Nondestructive detection of multiple qualities of dried jujube in different storage periods based on hyperspectral imaging combined with deep learning

Moisture content and sugar content are characteristic substances that determine the quality of dried jujubes. In this study, we collected visible/near-infrared and near-infrared hyperspectral data (Vis-NIR and NIR) from three dried jujubes samples with different storage periods, and established recognition models for different storage periods of jujubes using Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), Deep Learning Methods (LeNet, ResNet, DenseNet, MobileNet and EfficientNet). Then, we constructed a dried jujubes storage period classification model based on feature bands using Successive projections algorithm (SPA) and Principal component analysis (PCA) methods, the results showed that the classification results based on Vis-NIR were better than those based on NIR, and the classification model based on SPA method was more accurate. In addition, the internal quality attributes of moisture content and total sugar during a single storage period of jujube were also predicted, and a quality prediction model for multiple storage periods was established based on NIR. This study provides a fast and non-destructive method for detecting the storage and quality components of jujube, in order to control the quality of jujube at different storage periods.

Research on Atrial Fibrillation Diagnosis in Electrocardiograms Based on CLA-AF Model

Abstract Background The electrocardiogram (ECG) is the primary method for diagnosing atrial fibrillation (AF), but interpreting ECGs can be time-consuming and labor-intensive, which deserves more exploration. Methods We collected ECG data from 6,590 patients as YY2023, classified as Normal, AF, and Other. Convolutional Neural Network (CNN), bidirectional Long Short-Term Memory (BiLSTM), and Attention construct the AF recognition model CNN BiLSTM Attention-Atrial Fibrillation (CLA-AF). The generalization ability of the model is validated on public datasets CPSC2018, PhysioNet2017, and PTB-XL, and we explored the performance of oversampling, resampling, and hybrid datasets. Finally, additional PhysioNet2021 was added to validate the robustness and applicability in different clinical settings. We employed the SHAP method to interpret the model's predictions. Results The F1-score, Precision, and AUC of the CLA-AF model on YY2023 are 0.956, 0.970, and 1.00, respectively. Similarly, the AUC on CPSC2018, PhysioNet2017, and PTB-XL reached above 0.95, demonstrating its strong generalization ability. After oversampling PhysioNet2017, F1-score and Recall improved by 0.156 and 0.260. Generalization ability varied with sampling frequency. The model trained from the hybrid dataset has the most robust generalization ability, achieving an AUC of 0.96 or more. The AUC of PhysioNet2021 is 1.00, which proves the applicability of CLA-AF. The SHAP values visualization results demonstrate that the model's interpretation of AF aligns with the diagnostic criteria of AF. Conclusions The CLA-AF model demonstrates a high accuracy in recognizing AF from ECG, exhibiting remarkable applicability and robustness in diverse clinical settings.

Research on the Construction of Music Recognition Model Based on Improved CNN Convolutional Neural Network

Abstract. In the context of today's informationized society, music recognition technology within the field of intelligent audio processing has received extensive attention and become a research focus. However, traditional music recognition techniques are difficult to adapt to the complex changes of music signals due to the inherent defects in feature extraction and model construction. To address these difficulties, this study constructs an innovative music recognition algorithm, which is based on an optimized convolutional neural network (CNN). The efficacy of feature extraction is enhanced by fusing the hash convolutional neural network (Hash-CNN), and the temporal data is processed using the long short-term memory network (LSTM) to improve the accuracy of recognition. In the preprocessing stage of music signals, we applied various noise reduction and normalization means to enhance the data quality to a high standard. According to the experimental data, the model performs well in recognizing complex music segments, especially when analyzing multi-level and multi-dimensional music features, it shows strong robustness. Meanwhile, compared with traditional techniques, the model proposed in this study shows significant improvement in both recognition efficiency and accuracy, confirming its great potential and broad prospect in practical applications.

CTDNets: A High-Precision Hybrid Deep Learning Model for Modulation Recognition with Early-Stage Layer Fusion

To further enhance the recognition accuracy of automatic modulation recognition, improve communication efficiency, strengthen security, and optimize resource management, this paper designs a high-precision hybrid deep learning model featuring early-stage layer fusion. This model combines with Convolutional Neural Networks (CNN), Transformers, and Deep Neural Networks (DNN) to enhance the model’s feature extraction capabilities, thereby improving modulation recognition accuracy. Experiments are performed on RadioML2016.10a and RadioML2018.01a, and the results show that this architecture can effectively combine the advantages of different types of models, making the overall performance more robust and suitable for complex automatic modulation recognition problems.

GENERALIZED DATASET OF GEOLOGICAL AND GEOPHYSICAL INFORMATION ON THE EASTERN SECTOR OF THE RUSSIAN ARCTIC FOR MACHINE LEARNING – BASED ANALYSIS

The article presents a practical approach to the geological and geophysical spatial data collection and preliminary processing to use in machine learning models for geophysical applications. According to the established principles for estimating efforts in data analysis, which are confirmed by the results of surveys among specialists, this stage is viewed as major time and resource-consuming, amounting up to 80% in total volume of data analysis for a hypothesis testing project. The paper focuses on creating a consistent data set that integrates geological and geophysical information on a given region. We consider problems of different sources in the geodata representation to be related to their format (vector/raster), scale, type of attribute information (quantitative/qualitative) and their availability. The algorithm formalization and synthesis for combining geospatial data and converting them into quantitative vectors is a critical aspect. Combining various data draws on the concept of neighborhood fitting in with the data selection techniques and data consolidation strategy. The paper presents the general architecture of the software and hardware complex which includes a module for data collection and transformation in Python using the Pandas library, a data storage system based on the PostgreSQL DBMS (Database Management System) with the PostGIS extension. It is shown that for the considered class of problems in geophysics, it is sufficient to use a relational DBMS for data storing and processing. If the problem dimension increases, it is proposed to use the Big Data technology based on Apache Hadoop for scaling the system. A practical application of the proposed approach is demonstrated as results of data collection for the Caucasus region and eastern sector of the Russian Arctic. Based on the prepared data, experiments were carried out using machine learning models for recognition of locations of potential strong earthquakes and for sensitivity estimation of several geophysical features of these regions. The article presents the experimental results and evaluation of their efficiency.

Vocabulary-Enhanced Named Entity Recognition and its Application on Distribution Network Maintenance

Named entity recognition is a crucial task in natural language processing, serving downstream tasks such as information extraction and question-answering systems. With the success of Transformer in the field of natural language processing, named entity recognition has made significant strides. However, despite these advancements, the recognition capability in specialized domains remains challenging. The distribution network system contains a vast amount of equipment maintenance text, characterized by inherent complexity and specialization, lacking effective recognition and extraction methods. To address this issue, we propose a vocabulary-enhanced named entity recognition method. Unlike the original named entity recognition methods, we employ a vocabulary adapter to enhance the understanding of text semantic information by integrating matching information between the dictionary and the text. This method strengthens the ability to extract specialized nouns. In conjunction with the TaCL-BERT pre-trained model, we conducted training on our self-constructed distribution network maintenance dataset, GNP-M, to better align the model with the context and specific entities of the electrical grid maintenance domain. Simultaneously, we compared our method with four other named entity recognition models, revealing satisfactory performance results.

Improved facial emotion recognition model based on a novel deep convolutional structure

Facial Emotion Recognition (FER) is a very challenging task due to the varying nature of facial expressions, occlusions, illumination, pose variations, cultural and gender differences, and many other aspects that cause a drastic degradation in quality of facial images. In this paper, an anti-aliased deep convolution network (AA-DCN) model has been developed and proposed to explore how anti-aliasing can increase and improve recognition fidelity of facial emotions. The AA-DCN model detects eight distinct emotions from image data. Furthermore, their features have been extracted using the proposed model and numerous classical deep learning algorithms. The proposed AA-DCN model has been applied to three different datasets to evaluate its performance: The Cohn-Kanade Extending (CK+) database has been utilized, achieving an ultimate accuracy of 99.26% in (5 min, 25 s), the Japanese female facial expressions (JAFFE) obtained 98% accuracy in (8 min, 13 s), and on one of the most challenging FER datasets; the Real-world Affective Face (RAF) dataset; reached 82%, in low training time (12 min, 2s). The experimental results demonstrate that the anti-aliased DCN model is significantly increasing emotion recognition while improving the aliasing artifacts caused by the down-sampling layers.

Recognition and Spectral Analysis of Adi Digits

This article presents a method for automatically recognizing spoken digits in the Adi language using the highly effective open-source voice recognition toolkit, “Kaldi”. The study also includes an analysis of the formant frequencies and spectral characteristics of Adi digits. Adi is a zero-resource indigenous tribal language of Arunachal Pradesh with its roots in Tibeto-Burman. UNESCO's 2009 Atlas of the World's Languages in Danger classifies Adi as an endangered language in India. The research utilizes a modest digit corpus from 42 native Adi speakers, employing Mel Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP) features from continuous Adi speech utterances. The study explores both Monophone and Triphone models for digit recognition, with notable improvements in accuracy from 79.14% in the Monophone model to 84.29%, 88.43%, and 90.86% in the Triphone models (Tri1, Tri2, and Tri3). This proposed model holds promise for the development of Adi language speech recognition applications.

DeFFace: Deep Face Recognition Unlocked by Illumination Attributes

General face recognition is currently one of the key technologies in the field of computer vision, and it has achieved tremendous success with the support of deep-learning technology. General face recognition models currently exhibit extremely high accuracy on some high-quality face datasets. However, their performance decreases in challenging environments, such as low-light scenes. To enhance the performance of face recognition models in low-light scenarios, we propose a face recognition approach based on feature decoupling and fusion (DeFFace). Our main idea is to extract facial-related features from images that are not influenced by illumination. First, we introduce a feature decoupling network (D-Net) to decouple the image into facial-related features and illumination-related features. By incorporating the illumination triplet loss optimized with unpaired identity IDs, we regulate illumination-related features to minimize the impact of lighting conditions on the face recognition system. However, the decoupled features are relatively coarse. Therefore, we introduce a feature fusion network (F-Net) to further extract the residual facial-related features from the illumination-related features and fuse them with the initial facial-related features. Finally, we introduce a lighting-facial correlation loss to reduce the correlation between the two decoupled features in the specific space. We demonstrate the effectiveness of our method on four real-world low-light datasets and three simulated low-light datasets. We retrain multiple general face recognition methods using our proposed low-light training sets to further validate the advanced performance of our method. Compared to general face recognition methods, our approach achieves an average improvement of more than 2.11 percentage points on low-light face datasets. In comparison with image enhancement-based solutions, our method shows an average improvement of around 16 percentage points on low-light datasets, and it also delivers an average improvement of approximately 5.67 percentage points when compared to illumination normalization-based methods.

Skeleton-Based Human Action Recognition with Spatial and Temporal Attention-Enhanced Graph Convolution Networks

Skeleton-based human action recognition has great potential for human behavior analysis owing to its simplicity and robustness in varying environments. This paper presents a spatial and temporal attention-enhanced graph convolution network (STAEGCN) for human action recognition. The spatial-temporal attention module in the network uses convolution embedding for positional information and adopts multi-head self-attention mechanism to extract spatial and temporal attention separately from the input series of the skeleton. The spatial and temporal attention are then concatenated into an entire attention map according to a specific ratio. The proposed spatial and temporal attention module was integrated with an adaptive graph convolution network to form the backbone of STAEGCN. Based on STAEGCN, a two-stream skeleton-based human action recognition model was trained and evaluated. The model performed better on both NTU RGB+D and Kinetics 400 than 2s-AGCN and its variants. It was proven that the strategy of decoupling spatial and temporal attention and combining them in a flexible way helps improve the performance of graph convolution networks in skeleton-based human action recognition.

Automatic Modulation Recognition Based on Multimodal Information Processing: A New Approach and Application

Automatic modulation recognition (AMR) has wide applications in the fields of wireless communications, radar systems, and intelligent sensor networks. The existing deep learning-based modulation recognition models often focus on temporal features while overlooking the interrelations and spatio-temporal relationships among different types of signals. To overcome these limitations, a hybrid neural network based on a multimodal parallel structure, called the multimodal parallel hybrid neural network (MPHNN), is proposed to improve the recognition accuracy. The algorithm first preprocesses the data by parallelly processing the multimodal forms of the modulated signals before inputting them into the network. Subsequently, by combining Convolutional Neural Networks (CNN) and Bidirectional Gated Recurrent Unit (Bi-GRU) models, the CNN is used to extract spatial features of the received signals, while the Bi-GRU transmits previous state information of the time series to the current state to capture temporal features. Finally, the Convolutional Block Attention Module (CBAM) and Multi-Head Self-Attention (MHSA) are introduced as two attention mechanisms to handle the temporal and spatial correlations of the signals through an attention fusion mechanism, achieving the calibration of the signal feature maps. The effectiveness of this method is validated using various datasets, with the experimental results demonstrating that the proposed approach can fully utilize the information of multimodal signals. The experimental results show that the recognition accuracy of MPHNN on multiple datasets reaches 93.1%, and it has lower computational complexity and fewer parameters than other models.

Kernelized support vector machines for modified Gabor features facial recognition

The Gabor-filter approach has been extensively used in the recognition of patterns most especially in the extraction of features during image processing. Gabor filters usefulness explored in face recognition is traceable to its computational properties and biological relevance. Despite all the distinct characteristics of Gabor filters, it suffers from high feature dimensionality. This has led majorly to computational problems in any Gabor-based facial recognition model. The paper presents modified Gabor features for face recognition by introducing a meta-heuristics optimization algorithm using the Ant Colony Optimization Algorithm (ACO) to obtain relevant and optimal features from huge Gabor features. Kernels of Support Vector Machines (SVM); Linear SVM Kernel (LSVMK), Polynomial SVM Kernel (PSVMK), Sigmoid SVM (SSVMK) and Gaussian SVM Kernel (GSVMK) were employed for the classification of face images to either matched or mismatched. Two datasets were used for the evaluation of the system, they include: the Olivetti Research Laboratory (ORL) database and the acquired Africa face image database (ABFI). All the Kernelized SVMs produced an effective output in terms of training time, classification accuracy and error rate. Experimental results showed the lowest training time of 7.3195s was obtained in GSVMK for ABFI face image dataset, non-optimized Gabor feature gave the best accuracy of 90.88% in PSVMK of image size 75x75 for ABFI dataset, the optimized Gabor features recorded the best accuracy of 96.93% in GSVMK of image size 125x125 for ORL image dataset, the lowest error rate of 08.18% was obtained in LSVMK with image size of 150x150 for ORL image dataset

Using Large Language Models to Abstract Complex Social Determinants of Health From Original and Deidentified Medical Notes: Development and Validation Study.

Social determinants of health (SDoH) such as housing insecurity are known to be intricately linked to patients' health status. More efficient methods for abstracting structured data on SDoH can help accelerate the inclusion of exposome variables in biomedical research and support health care systems in identifying patients who could benefit from proactive outreach. Large language models (LLMs) developed from Generative Pre-trained Transformers (GPTs) have shown potential for performing complex abstraction tasks on unstructured clinical notes. Here, we assess the performance of GPTs on identifying temporal aspects of housing insecurity and compare results between both original and deidentified notes. We compared the ability of GPT-3.5 and GPT-4 to identify instances of both current and past housing instability, as well as general housing status, from 25,217 notes from 795 pregnant women. Results were compared with manual abstraction, a named entity recognition model, and regular expressions. Compared with GPT-3.5 and the named entity recognition model, GPT-4 had the highest performance and had a much higher recall (0.924) than human abstractors (0.702) in identifying patients experiencing current or past housing instability, although precision was lower (0.850) compared with human abstractors (0.971). GPT-4's precision improved slightly (0.936 original, 0.939 deidentified) on deidentified versions of the same notes, while recall dropped (0.781 original, 0.704 deidentified). This work demonstrates that while manual abstraction is likely to yield slightly more accurate results overall, LLMs can provide a scalable, cost-effective solution with the advantage of greater recall. This could support semiautomated abstraction, but given the potential risk for harm, human review would be essential before using results for any patient engagement or care decisions. Furthermore, recall was lower when notes were deidentified prior to LLM abstraction.

A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities.

Biomedical image analysis is fundamental for biomedical discovery. Holistic image analysis comprises interdependent subtasks such as segmentation, detection and recognition, which are tackled separately by traditional approaches. Here, we propose BiomedParse, a biomedical foundation model that can jointly conduct segmentation, detection and recognition across nine imaging modalities. This joint learning improves the accuracy for individual tasks and enables new applications such as segmenting all relevant objects in an image through a textual description. To train BiomedParse, we created a large dataset comprising over 6 million triples of image, segmentation mask and textual description by leveraging natural language labels or descriptions accompanying existing datasets. We showed that BiomedParse outperformed existing methods on image segmentation across nine imaging modalities, with larger improvement on objects with irregular shapes. We further showed that BiomedParse can simultaneously segment and label all objects in an image. In summary, BiomedParse is an all-in-one tool for biomedical image analysis on all major image modalities, paving the path for efficient and accurate image-based biomedical discovery.

Toward Improving Human Training by Combining Wearable Full-Body IoT Sensors and Machine Learning.

This paper proposes DigitalUpSkilling, a novel IoT- and AI-based framework for improving and personalising the training of workers who are involved in physical-labour-intensive jobs. DigitalUpSkilling uses wearable IoT sensors to observe how individuals perform work activities. Such sensor observations are continuously processed to synthesise an avatar-like kinematic model for each worker who is being trained, referred to as the worker's digital twins. The framework incorporates novel work activity recognition using generative adversarial network (GAN) and machine learning (ML) models for recognising the types and sequences of work activities by analysing an individual's kinematic model. Finally, the development of skill proficiency ML is proposed to evaluate each trainee's proficiency in work activities and the overall task. To illustrate DigitalUpSkilling from wearable IoT-sensor-driven kinematic models to GAN-ML models for work activity recognition and skill proficiency assessment, the paper presents a comprehensive study on how specific meat processing activities in a real-world work environment can be recognised and assessed. In the study, DigitalUpSkilling achieved 99% accuracy in recognising specific work activities performed by meat workers. The study also presents an evaluation of the proficiency of workers by comparing kinematic data from trainees performing work activities. The proposed DigitalUpSkilling framework lays the foundation for next-generation digital personalised training.