Deep Learning Algorithms Research Articles

Tactical intent-driven autonomous air combat behavior generation method

With the rapid development and deep application of artificial intelligence, modern air combat is incrementally evolving towards intelligent combat. Although deep reinforcement learning algorithms have contributed to dramatic advances in in air combat, they still face challenges such as poor interpretability and weak transferability of adversarial strategies. In this regard, this paper proposes a tactical intent-driven method for autonomous air combat behaviour generation. Firstly, this paper explores the mapping relationship between optimal strategies and rewards, demonstrating the detrimental effects of the combination of sparse rewards and dense rewards on policy. Built around this, the decision-making process of pilot behavior is analyzed, and a reward mapping model from intent to behavior is established. Finally, to address the problems of poor stability and slow convergence speed of deep reinforcement learning algorithms in large-scale state-action spaces, the dueling-noisy-multi-step DQN algorithm is devised, which not only improves the accuracy of value function approximation but also enhances the efficiency of space exploration and network generalization. Through experiments, the conflicts between sparse rewards and dense rewards are demonstrated. The superior performance and stability of the proposed algorithm compared to other algorithms are captured by our empirical results. More intuitively, the strategies under different intents exhibit strong interpretability and flexibility, which can provide tactical support for intelligent decision-making in air combat.

Hybrid LSTM–Attention and CNN Model for Enhanced Speech Emotion Recognition

Emotion recognition is crucial for enhancing human–machine interactions by establishing a foundation for AI systems that integrate cognitive and emotional understanding, bridging the gap between machine functions and human emotions. Even though deep learning algorithms are actively used in this field, the study of sequence modeling that accounts for the shifts in emotions over time has not been thoroughly explored. In this research, we present a comprehensive speech emotion-recognition framework that amalgamates the ZCR, RMS, and MFCC feature sets. Our approach employs both CNN and LSTM networks, complemented by an attention model, for enhanced emotion prediction. Specifically, the LSTM model addresses the challenges of long-term dependencies, enabling the system to factor in historical emotional experiences alongside current ones. We also incorporate the psychological “peak–end rule”, suggesting that preceding emotional states significantly influence the present emotion. The CNN plays a pivotal role in restructuring input dimensions, facilitating nuanced feature processing. We rigorously evaluated the proposed model utilizing two distinct datasets, namely TESS and RAVDESS. The empirical outcomes highlighted the model’s superior performance, with accuracy rates reaching 99.8% for TESS and 95.7% for RAVDESS. These results are a notable advancement, showcasing our system’s precision and innovative contributions to emotion recognition.

Satellite-Based and Street-View Green Space and Adiposity in US Children

Green space may be associated with obesity outcomes in children by promoting physical activity and reducing stress. However, previous studies have mostly used a single measure of green space and body mass index (BMI) as the sole indicator of adiposity. To examine associations of the satellite-based normalized difference vegetation index (NDVI) and street-view imagery-based green space metrics with comprehensive adiposity measures in children. The Project Viva cohort study observed participants from mid-childhood (mean [SD] age, 8.0 [0.8] years) and early adolescence (mean [SD] age, 13.3 [1.0] years) through late adolescence (mean [SD] age, 17.8 [0.7] years) between January 1, 2007, and December 31, 2021. This analysis included children with data on green space and adiposity measures. Data were analyzed from January 1 to June 30, 2024. Green space metrics (ie, percentage of trees, grass, and other green space) using street-view imagery and deep learning algorithms along with the NDVI were linked with participants' home addresses during mid-childhood and early adolescence. Body mass index (calculated as weight in kilograms divided by height in meters squared) and waist circumference z scores were calculated from anthropometric measurements during early and late adolescence. The total fat mass index (FMI; calculated as fat mass in kilograms divided by height in meters squared) and trunk FMI were obtained using dual-energy x-ray absorptiometry. Separate linear regression models adjusted for a child's age, sex, individual and neighborhood socioeconomic status, and population density estimated prospective associations of the NDVI and street-view green space (in mid-childhood or early adolescence) with 4 adiposity measures (in early or late adolescence). The sample included 843 children (mean [SD] age at mid-childhood, 7.9 [0.8] years; 423 girls [50.2%]) at baseline. After adjustment for confounders, a higher NDVI (per 1-SD increase) within 270 m in early adolescence was associated with lower z scores for BMI (-0.16 [95% CI, -0.27 to -0.04]) and waist circumference (-0.13 [95% CI, -0.23 to -0.03]), as well as a lower total FMI (-0.62 [95% CI, -1.02 to -0.23]) and trunk FMI (-0.33 [95% CI, -0.55 to -0.12]) in late adolescence. Higher percentages of street-view other green space (flowers, plants, and fields), but not street-view trees and grass, within 500 m in early adolescence were associated with lower BMI z scores (-0.09 [95% CI, -0.18 to -0.01) in late adolescence. The NDVI and street-view green space exposure at mid-childhood were not associated with adiposity measures in early adolescence. This cohort study of US children suggests that higher levels of satellite-based NDVI greenness and percentages of street-level green space components (flowers, plants, and fields) were associated with lower adiposity. The findings support the exploration of increasing residential green space levels and adding specific green space components as an urban planning and public health intervention strategy to combat the prevalence of childhood obesity in the US.

High-dimensional anticounterfeiting nanodiamonds authenticated with deep metric learning

Physical unclonable function labels have emerged as a promising candidate for achieving unbreakable anticounterfeiting. Despite their significant progress, two challenges for developing practical physical unclonable function systems remain, namely 1) fairly few high-dimensional encoded labels with excellent material properties, and 2) existing authentication methods with poor noise tolerance or inapplicability to unseen labels. Herein, we employ the linear polarization modulation of randomly distributed fluorescent nanodiamonds to demonstrate, for the first time, three-dimensional encoding for diamond-based labels. Briefly, our three-dimensional encoding scheme provides digitized images with an encoding capacity of 109771 and high distinguishability under a short readout time of 7.5 s. The high photostability and inertness of fluorescent nanodiamonds endow our labels with high reproducibility and long-term stability. To address the second challenge, we employ a deep metric learning algorithm to develop an authentication methodology that computes the similarity of deep features of digitized images, exhibiting a better noise tolerance than the classical point-by-point comparison method. Meanwhile, it overcomes the key limitation of existing artificial intelligence-driven classification-based methods, i.e., inapplicability to unseen labels. Considering the high performance of both fluorescent nanodiamonds labels and deep metric learning authentication, our work provides the basis for developing practical physical unclonable function anticounterfeiting systems.

Z-Score Experience Replay in Off-Policy Deep Reinforcement Learning

Reinforcement learning, as a machine learning method that does not require pre-training data, seeks the optimal policy through the continuous interaction between an agent and its environment. It is an important approach to solving sequential decision-making problems. By combining it with deep learning, deep reinforcement learning possesses powerful perception and decision-making capabilities and has been widely applied to various domains to tackle complex decision problems. Off-policy reinforcement learning separates exploration and exploitation by storing and replaying interaction experiences, making it easier to find global optimal solutions. Understanding how to utilize experiences is crucial for improving the efficiency of off-policy reinforcement learning algorithms. To address this problem, this paper proposes Z-Score Prioritized Experience Replay, which enhances the utilization of experiences and improves the performance and convergence speed of the algorithm. A series of ablation experiments demonstrate that the proposed method significantly improves the effectiveness of deep reinforcement learning algorithms.

基于改进轻量化卷积神经网络模型的玉露香梨缺陷识别研究

The skin of Yu-Lu-Xiang pears is brittle and easily damaged during picking and sorting. In order to reduce the secondary damage caused by mechanical automatic sorting of Yu-Lu-Xiang pear after harvest, optimize the sorting process and improve the sorting accuracy. Based on the MobileV2Net model, a lightweight convolutional neural network model EC-MobileV2Net-Fast, which integrated transfer learning and attention mechanism, was proposed to identify skin damage defects of Yu-Lu-Xiang pears. According to the defects of Yu-Lu-Xiang pears with different damage degrees, a dataset containing four characteristics was created. The model accuracy rate, single defect identification accuracy rate, recall, specificity, parameter and so on were taken as evaluation indexes, and the interpretation ability of the model decision was analyzed by Grad-CAM thermal map. Preliminary evaluation results showed that the model produced the highest level of accuracy, underscoring the potential of deep learning algorithms to significantly enhance defect recognition and classification. It can improve sorting efficiency, reduce labor costs and strictly control after-sales quality.

An Automatic Modulation Recognition Algorithm Based on Time–Frequency Features and Deep Learning with Fading Channels

Automatic modulation recognition (AMR) stands as a crucial core technology within the realm of signal processing and perception, playing a significant part in harsh electromagnetic environments. The time–frequency image (TFI) of communication signals can manifest modulation characteristics and serve as a foundation for signal modulation recognition and classification. However, under the influence of the electromagnetic environment, communication signals are exposed to varying degrees of interference, which poses a challenge to the recognition of modulation types. Taking into account the effects of interference and channel fading, this paper introduces a communication signal modulation recognition algorithm based on deep learning (DL) and time–frequency analysis. This approach employs short-time Fourier transform (STFT) to generate time–frequency diagrams from time-domain signals. Subsequently, it binarizes the image and feeds it as input data to the neural network. Our research presents a composite deep convolutional neural network (CNN) architecture known as the composite dense-residual neural network (CDRNN). This architecture focuses on enhancing the feature extraction and identification, aiming to achieve accurate recognition of modulation types in harsh electromagnetic environments. Finally, simulation results validate that the proposed deep learning algorithm holds remarkable advantages in boosting the accuracy of modulation type recognition with better adaptability. The algorithm shows better performance even in harsh electromagnetic environments. When the signal-to-noise ratio (SNR) is 18 dB, the recognition accuracy can reach 92.1%.

A novel deep learning algorithm for real-time prediction of clinical deterioration in the emergency department for a multimodal clinical decision support system.

The array of complex and evolving patient data has limited clinical decision making in the emergency department (ED). This study introduces an advanced deep learning algorithm designed to enhance real-time prediction accuracy for integration into a novel Clinical Decision Support System (CDSS). A retrospective study was conducted using data from a level 1 tertiary hospital. The algorithm's predictive performance was evaluated based on in-hospital cardiac arrest, inotropic circulatory support, advanced airway, and intensive care unit admission. We developed an artificial intelligence (AI) algorithm for CDSS that integrates multiple data modalities, including vitals, laboratory, and imaging results from electronic health records. The AI model was trained and tested on a dataset of 237,059 ED visits. The algorithm's predictions, based solely on triage information, significantly outperformed traditional logistic regression models, with notable improvements in the area under the precision-recall curve (AUPRC). Additionally, predictive accuracy improved with the inclusion of continuous data input at shorter intervals. This study suggests the feasibility of using AI algorithms in diverse clinical scenarios, particularly for earlier detection of clinical deterioration. Future work should focus on expanding the dataset and enhancing real-time data integration across multiple centers to further optimize its application within the novel CDSS.

Morphologies of galaxies within voids

Among the largest structures in which matter is distributed in the Universe, we find cosmic voids, which are large, under-dense regions almost devoid of galaxies. The study of these structures and the galaxies that inhabit them, the void galaxies, provides key information for understanding galaxy evolution. In this work we investigate the effects of the environment on the evolution of void galaxies. In particular, we study their morphology and explore its dependence on the location within the void where the galaxies reside, as well as on the properties of the void, such as its size and the galaxy number density. The sample of void galaxies that we use in this study is based on the catalogue of cosmic voids and void galaxies in the Sloan Digital Sky Survey Data Release 7 (SDSS-DR7). As we are interested in studying the morphology of void galaxies, we select galaxies in the redshift range of 0.005\,leq \,z\,leq \,0.080, and use the public galaxy morphologies of the SDSS sample together with deep learning algorithms to divide the sample into early- and late-type void galaxies. We analyse the fractions of galaxies of each morphological type as a function of the void-centric distance, the size of the voids, and the density of galaxies in each void. There is a higher abundance of late-type galaxies with respect to early-type galaxies within voids, which remains nearly constant from the inner to the outer part of the voids. We do not find any dependence of the fraction of early- and late-type galaxies on void size or on the number-density of galaxies in the voids. Galaxies in voids follow the morphology--density relation, in the sense that the majority of the galaxies in voids (the most under-dense large-scale environments) are late-type galaxies. However, we find no difference between voids with lower or higher volume number-density of galaxies: the fractions of early- and late-type galaxies do not depend on the density of the voids. The physical processes responsible for the evolution from late towards earlier types (such as external environmental quenching) are not sufficiently effective in voids or are so slow (internal secular quenching) that their contributions do not appear in the morphology--density relation.

Development of an Open-Source Dataset of Flat-Mounted Images for the Murine Oxygen-Induced Retinopathy Model of Ischemic Retinopathy.

To describe an open-source dataset of flat-mounted retinal images and vessel segmentations from mice subject to the oxygen-induced retinopathy (OIR) model. Flat-mounted retinal images from mice killed at postnatal days 12 (P12), P17, and P25 used in prior OIR studies were compiled. Mice subjected to normoxic conditions were killed at P12, P17, and P25, and their retinas were flat-mounted for imaging. Major blood vessels from the OIR images were manually segmented by four graders (JSC, HKR, KBL, JM), with cross-validation performed to ensure similar grading. Overall, 1170 images were included in this dataset. Of these images, 111 were of normoxic mice retina, and 1048 were mice subject to OIR. The majority of images from OIR mice were obtained at P17. The 50 images obtained from an external dataset, OIRSeg, did not have age labels. All images were manually segmented and used in the training or testing of a previously published deep learning algorithm. This is the first open-source dataset of original and segmented flat-mounted retinal images. The dataset has potential applications for expanding the development of generalizable and larger-scale artificial intelligence and analyses for OIR. This dataset is published online and publicly available at dx.doi.org/10.6084/m9.figshare.23690973. This open access dataset serves as a source of raw data for future research involving big data and artificial intelligence research concerning oxygen-induced retinopathy.

Development and validation of a deep learning algorithm for the classification of the level of surgical difficulty in impacted mandibular third molar surgery

The aim of this study was to develop and validate a convolutional neural network (CNN) algorithm for the detection of impacted mandibular third molars in panoramic radiographs and the classification of the surgical extraction difficulty level. A dataset of 1730 panoramic radiographs was collected; 1300 images were allocated to training and 430 to testing. The performance of the model was evaluated using the confusion matrix for multiclass classification, and the actual scores were compared to those of two human experts. The area under the precision–recall curve of the YOLOv5 model ranged from 72% to 89% for the variables in the surgical difficulty index (level of occlusion, available retromolar space, impaction angle, number of roots, root curvature, and relationship to the second molar). The area under the receiver operating characteristic curve showed promising results of the YOLOv5 model for classifying third molars into three surgical difficulty levels (micro-average AUC 87%). Furthermore, the algorithm scores demonstrated good agreement with the human experts. In conclusion, the YOLOv5 model has the potential to accurately detect and classify the position of mandibular third molars, with high performance for every criterion in radiographic images. The proposed model could serve as an aid in improving clinician performance and could be integrated into a screening system.

DLLabelsCT: Annotation tool using deep transfer learning to assist in creating new datasets from abdominal computed tomography scans, case study: Pancreas.

The utilization of artificial intelligence (AI) is expanding significantly within medical research and, to some extent, in clinical practice. Deep learning (DL) applications, which use large convolutional neural networks (CNN), hold considerable potential, especially in optimizing radiological evaluations. However, training DL algorithms to clinical standards requires extensive datasets, and their processing is labor-intensive. In this study, we developed an annotation tool named DLLabelsCT that utilizes CNN models to accelerate the image analysis process. To validate DLLabelsCT, we trained a CNN model with a ResNet34 encoder and a UNet decoder to segment the pancreas on an open-access dataset and used the DL model to assist in annotating a local dataset, which was further used to refine the model. DLLabelsCT was also tested on two external testing datasets. The tool accelerates annotation by 3.4 times compared to a completely manual annotation method. Out of 3,715 CT scan slices in the testing datasets, 50% did not require editing when reviewing the segmentations made by the ResNet34-UNet model, and the mean and standard deviation of the Dice similarity coefficient was 0.82±0.24. DLLabelsCT is highly accurate and significantly saves time and resources. Furthermore, it can be easily modified to support other deep learning models for other organs, making it an efficient tool for future research involving larger datasets.

Evaluating Performances of LSTM, SVM, GPR, and RF for Drought Prediction in Norway: A Wavelet Decomposition Approach on Regional Forecasting

A serious natural disaster that poses a threat to people and their living spaces is drought, which is difficult to notice at first and can quickly spread to wide areas through subtle progression. Numerous methods are being explored to identify, prevent, and mitigate drought, and distinct metrics have been developed. In order to contribute to the research on measures to be taken against drought, the Standard Precipitation Evaporation Index (SPEI), one of the drought indices that has been developed and accepted in recent years and includes a more comprehensive drought definition, was chosen in this study. Machine learning and deep learning algorithms, including support vector machine (SVM), random forest (RF), long short-term memory (LSTM), and Gaussian process regression (GPR), were used to model the droughts in six regions of Norway: Bodø, Karasjok, Oslo, Tromsø, Trondheim, and Vadsø. Four distinct model architectures were employed for this goal, and as a novel approach, the models’ output was enhanced by using discrete wavelet decomposition/transformation (WT). The model outputs were evaluated using the correlation coefficient (r), Nash–Sutcliffe efficiency (NSE), and root mean square error (RMSE) as performance evaluation criteria. When the findings were analyzed, the GPR model (W-GPR), which was acquired after WT, typically produced the best results. Furthermore, it was discovered that, out of all the recognized models, M04 had the most effective model structure. Consequently, the most successful outcomes were obtained with W-SVM-M04 for Bodø and W-GPR-M04 for Karasjok, Oslo, Tromsø, Trondheim, and Vadsø. Furthermore, W-GPR-M04 in the Oslo region had the best results across all regions (r: 0.9983, NSE: 0.9966 and RMSE:0.0539).

Investigation and application of big data and AI-Powered provincial power grid accurate planning support platform

ABSTRACT This study presents an accurate planning assistance platform for big data and artificial intelligence-based provincial power grids. The platform greatly increases the efficiency of data integration and management by building a planning big database containing fundamental geographic information, thematic data, remote sensing image data, and planning business data. Technically speaking, this paper addresses the issues of accuracy and speed in the processing of remote sensing image data by using deep learning algorithms like convolutional neural networks (CNN) and full convolutional neural networks (FCN) to achieve automatic classification of high-resolution optical remote sensing images and residential area extraction. In the meantime, the intelligent routing approach, which is based on cost surface modelling and multi-criteria decision-making, optimises the routing of power lines by carefully taking into account a wide range of variables, including land use, accessibility for vehicles, and environmental impact. In addition to resolving the primary issues with present power grid planning, this study’s exact planning support platform for provincial power grids greatly enhances real-time, accurate, and scientific planning.

Multimodal MRI Deep Learning for Predicting Central Lymph Node Metastasis in Papillary Thyroid Cancer

Background: Central lymph node metastasis (CLNM) in papillary thyroid cancer (PTC) significantly influences surgical decision-making strategies. Objectives: This study aims to develop a predictive model for CLNM in PTC patients using magnetic resonance imaging (MRI) and clinicopathological data. Methods: By incorporating deep learning (DL) algorithms, the model seeks to address the challenges in diagnosing CLNM and reduce overtreatment. The results were compared with traditional machine learning (ML) models. In this retrospective study, preoperative MRI data from 105 PTC patients were divided into training and testing sets. A radiologist manually outlined the region of interest (ROI) on MRI images. Three classic ML algorithms (support vector machine [SVM], logistic regression [LR], and random forest [RF]) were employed across different data modalities. Additionally, an AMMCNet utilizing convolutional neural networks (CNNs) was proposed to develop DL models for CLNM. Predictive performance was evaluated using receiver operator characteristic (ROC) curve analysis, and clinical utility was assessed through decision curve analysis (DCA). Results: Lesion diameter was identified as an independent risk factor for CLNM. Among ML models, the RF-(T1WI + T2WI, T1WI + T2WI + Clinical) models achieved the highest area under the curve (AUC) at 0.863. The DL fusion model surpassed all ML fusion models with an AUC of 0.891. Conclusions: A fusion model based on the AMMCNet architecture using MRI images and clinicopathological data was developed, effectively predicting CLNM in PTC patients.

Optimizing Stroke Classification with Pre-Trained Deep Learning Models

Background/Objectives: Insufficient blood supply to the brain, whether due to blocked arteries (ischemic stroke) or bleeding (hemorrhagic stroke), leads to brain cell death and cognitive impairment. Ischemic strokes, which are more common, occur when blood flow to the brain is obstructed. Magnetic resonance imaging (MRI) scans are essential for distinguishing stroke types, but precise and timely identification of ischemic strokes is crucial for effective treatment. Manual diagnosis can be difficult due to high patient volumes and time constraints in hospitals. This study aims to investigate the use of deep learning techniques for predicting ischemic strokes with high accuracy, enabling earlier diagnosis and intervention. Methods: The study utilized advanced deep learning algorithms, specifically ConvNeXt Base, to analyze large datasets of medical imaging data, focusing on MRI scans. The model was trained and validated on a labeled dataset to identify critical indicators and patterns associated with stroke risk. The performance of the model was evaluated based on accuracy metrics to determine its predictive capabilities. Results: ConvNeXt Base achieved an overall accuracy of 84% on the validation set, demonstrating its effectiveness in identifying ischemic strokes. The model was able to detect key patterns linked to stroke risk, highlighting its potential for use in clinical settings to aid in early diagnosis and decision-making. Conclusions: ConvNeXt Base reveals promise in improving stroke prediction accuracy, enabling earlier diagnosis and personalized treatment, which could lead to faster, more effective medical interventions.

Deep Learning Algorithm for Optimized Sensor Data Fusion in Fault Diagnosis and Tolerance

Environmental perception is one of the key technologies to realize autonomous vehicles. The fault diagnosis process involves identifying the fault that occurred or the cause of the out-of-control condition. Here, the major objective is to locate problems in detection by analysing previous data or sequential patterns of data that cause failure. This study evaluates the use of deep learning for improved sensor data fusion in fault identification and tolerance using the KITTI dataset. The input video from the dataset has been transformed to frames through median filtering. Next, feature extraction is applied to a preprocessed image, resulting in the fusion of sensor data. Data fusion is then carried out utilizing an enhanced RPN (region proposal network). The enhanced RPN also has a loss function (object detection loss, bounding box loss and target classification loss), an estimate of ROI and feature extraction network (FEN). Through the use of the COOT connected blue monkey optimization (CCBMO) model, the weight of the optimally enhanced RPN is established. Next, using global non-maximum suppression with both global and local confidence, fault identification and tolerance are carried out. From the analysis, it clearly shows that proposed method accomplished better results in terms of accuracy, precision and specificity of 97.78%, 93.76% and 93.43%, respectively, when compared with various conventional models with respect to diverse performance measures.

Recognition of Radix Bupleuri origin using laser-induced breakdown spectroscopy (LIBS) combined with deep learning and machine learning algorithms

In this study, laser-induced breakdown spectroscopy (LIBS) was utilized to classify Radix Bupleuri from six origins to acquire good results. Several feature variable selection methods have been used to preprocess the experimental data on the LIBS. Machine learning and deep learning algorithms were used to build classification models, respectively. The accuracy of all four classification models on the prediction set is higher than 99%. Among them, the classification model constructed by the residual neural network (ResNet) algorithm demonstrated the optimal prediction accuracy both in the full waveform and feature variable selection. For the independent prediction set, the ResNet classification model still demonstrated optimal prediction with 88.33% accuracy. On the importance ranking of feature variables, the elements as Cl, Mg, Fe, Ba, Li, O, and He were considered to play an important role in the identification of the origin of Radix Bupleuri. This study demonstrated that LIBS combined with deep learning and machine learning algorithms can provide a more efficient and reliable method for the origin identification of Radix Bupleuri. In addition, it can be found that the application of deep learning algorithms can realize the prediction output more efficiently and accurately for the analysis of high-dimensional data of LIBS.

An artificial intelligence approach to smart exam supervision using YOLOv5 and siamese network

<p>Artificial intelligence has introduced revolutionary and innovative solutions to many complex problems by automating processes or tasks that used to require human power. The limited capabilities of human efforts in real-time monitoring have led to artificial intelligence becoming increasingly popular. Artificial intelligence helps develop the monitoring process by analysing data and extracting accurate results. Artificial intelligence is also capable of providing surveillance cameras with a digital brain that analyses images and live video clips without human intervention. Deep learning models can be applied to digital images to identify and classify objects accurately. Object detection algorithms are based on deep learning algorithms in artificial intelligence. Using the deep learning algorithm, object detection is achieved with high accuracy. In this paper, a combined model of the YOLOv5 model and network Siames technology is proposed, in which the YOLOv5 algorithm detects cheating tools in classrooms, such as a cell phone or a book, in such away that the algorithm detects the student as an object and cannot recognize his face. Using the Siames network, we compare the student’s face against the data base of students in order to identify the student with cheating tools.</p>

Detection of black tea fermentation quality based on optimized deep neural network and hyperspectral imaging

The quality of black tea significantly relies on its fermentation process. Nevertheless, achieving precise and objective evaluations remains challenging due to the subjective nature of manual judgment involved in quality monitoring. To address this problem, hyperspectral imaging combined with the deep learning algorithms are proposed to identify the fermentation quality of black tea. Firstly, the hyperspectral data of Yinghong No. 9 black tea during five fermentation time intervals within 0–5 h are collected. Then, the Support Vector Machine (SVM), Artificial Neural Network (ANN), Partial Least Squares Discriminant Analysis (PLS-DA), and Naive Bayesian (NB) are used to construct black tea fermentation quality detection models based on full spectrum and selected spectrum data. Furthermore, deep learning algorithms including the Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), CNN-LSTM, and Swarm Optimization (PSO) optimized CNN-LSTM (PSO-CNN-LSTM) are also used to build the detection model using the spectral images. The experimental results indicate that deep learning algorithms have obvious advantages over traditional machine learning algorithms in tea fermentation quality detection. Besides, the PSO-CNN-LSTM model shows the best classification performance compared to other algorithms and achieves an accuracy of 96.78% on the test set. This study demonstrates the significant potential of combining deep learning with hyperspectral imaging for predicting black tea fermentation quality. This provides a new approach for effective monitoring of the black tea fermentation process and a useful reference for other applications in similar fields.