An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect s...An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect sample data for light guide plate(LGP) in production,as well as the problem of minor defects.Two optimizations are made to the generator of CycleGAN:fusion of low resolution features obtained from partial up-sampling and down-sampling with high-resolution features,combination of self attention mechanism with residual network structure to replace the original residual module.Qualitative and quantitative experiments were conducted to compare different data augmentation methods,and the results show that the defect images of the LGP generated by the improved network were more realistic,and the accuracy of the you only look once version 5(YOLOv5) detection network for the LGP was improved by 5.6%,proving the effectiveness and accuracy of the proposed method.展开更多
The integration of image analysis through deep learning(DL)into rock classification represents a significant leap forward in geological research.While traditional methods remain invaluable for their expertise and hist...The integration of image analysis through deep learning(DL)into rock classification represents a significant leap forward in geological research.While traditional methods remain invaluable for their expertise and historical context,DL offers a powerful complement by enhancing the speed,objectivity,and precision of the classification process.This research explores the significance of image data augmentation techniques in optimizing the performance of convolutional neural networks(CNNs)for geological image analysis,particularly in the classification of igneous,metamorphic,and sedimentary rock types from rock thin section(RTS)images.This study primarily focuses on classic image augmentation techniques and evaluates their impact on model accuracy and precision.Results demonstrate that augmentation techniques like Equalize significantly enhance the model's classification capabilities,achieving an F1-Score of 0.9869 for igneous rocks,0.9884 for metamorphic rocks,and 0.9929 for sedimentary rocks,representing improvements compared to the baseline original results.Moreover,the weighted average F1-Score across all classes and techniques is 0.9886,indicating an enhancement.Conversely,methods like Distort lead to decreased accuracy and F1-Score,with an F1-Score of 0.949 for igneous rocks,0.954 for metamorphic rocks,and 0.9416 for sedimentary rocks,exacerbating the performance compared to the baseline.The study underscores the practicality of image data augmentation in geological image classification and advocates for the adoption of DL methods in this domain for automation and improved results.The findings of this study can benefit various fields,including remote sensing,mineral exploration,and environmental monitoring,by enhancing the accuracy of geological image analysis both for scientific research and industrial applications.展开更多
Data augmentation plays an important role in boosting the performance of 3D models,while very few studies handle the 3D point cloud data with this technique.Global augmentation and cut-paste are commonly used augmenta...Data augmentation plays an important role in boosting the performance of 3D models,while very few studies handle the 3D point cloud data with this technique.Global augmentation and cut-paste are commonly used augmentation techniques for point clouds,where global augmentation is applied to the entire point cloud of the scene,and cut-paste samples objects from other frames into the current frame.Both types of data augmentation can improve performance,but the cut-paste technique cannot effectively deal with the occlusion relationship between the foreground object and the background scene and the rationality of object sampling,which may be counterproductive and may hurt the overall performance.In addition,LiDAR is susceptible to signal loss,external occlusion,extreme weather and other factors,which can easily cause object shape changes,while global augmentation and cut-paste cannot effectively enhance the robustness of the model.To this end,we propose Syn-Aug,a synchronous data augmentation framework for LiDAR-based 3D object detection.Specifically,we first propose a novel rendering-based object augmentation technique(Ren-Aug)to enrich training data while enhancing scene realism.Second,we propose a local augmentation technique(Local-Aug)to generate local noise by rotating and scaling objects in the scene while avoiding collisions,which can improve generalisation performance.Finally,we make full use of the structural information of 3D labels to make the model more robust by randomly changing the geometry of objects in the training frames.We verify the proposed framework with four different types of 3D object detectors.Experimental results show that our proposed Syn-Aug significantly improves the performance of various 3D object detectors in the KITTI and nuScenes datasets,proving the effectiveness and generality of Syn-Aug.On KITTI,four different types of baseline models using Syn-Aug improved mAP by 0.89%,1.35%,1.61%and 1.14%respectively.On nuScenes,four different types of baseline models using Syn-Aug improved mAP by 14.93%,10.42%,8.47%and 6.81%respectively.The code is available at https://github.com/liuhuaijjin/Syn-Aug.展开更多
Bird species classification is not only a challenging topic in artificial intelligence but also a domain closely related to environmental protection and ecological research.Additionally,performing edge computing on lo...Bird species classification is not only a challenging topic in artificial intelligence but also a domain closely related to environmental protection and ecological research.Additionally,performing edge computing on low-level devices using small neural networks can be an important research direction.In this paper,we use the EfficientNetV2B0 model for bird species classification,applying transfer learning on a dataset of 525 bird species.We also employ the BiRefNet model to remove backgrounds from images in the training set.The generated background-removed images are mixed with the original training set as a form of data augmentation.We aim for these background-removed images to help the model focus on key features,and by combining data augmentation with transfer learning,we trained a highly accurate and efficient bird species classification model.The training process is divided into a transfer learning stage and a fine-tuning stage.In the transfer learning stage,only the newly added custom layers are trained;while in the fine-tuning stage,all pre-trained layers except for the batch normalization layers are fine-tuned.According to the experimental results,the proposed model not only has an advantage in size compared to other models but also outperforms them in various metrics.The training results show that the proposed model achieved an accuracy of 99.54%and a precision of 99.62%,demonstrating that it achieves both lightweight design and high accuracy.To confirm the credibility of the results,we use heatmaps to interpret the model.The heatmaps show that our model can clearly highlight the image feature area.In addition,we also perform the 10-fold cross-validation on the model to verify its credibility.Finally,this paper proposes a model with low training cost and high accuracy,making it suitable for deployment on edge computing devices to provide lighter and more convenient services.展开更多
In order to address the widespread data shortage problem in battery research,this paper proposes a generative adversarial network model that combines it with deep convolutional networks,the Wasserstein distance,and th...In order to address the widespread data shortage problem in battery research,this paper proposes a generative adversarial network model that combines it with deep convolutional networks,the Wasserstein distance,and the gradient penalty to achieve data augmentation.To lower the threshold for implementing the proposed method,transfer learning is further introduced.The W-DC-GAN-GP-TL framework is thereby formed.This framework is evaluated on 3 different publicly available datasets to judge the quality of generated data.Through visual comparisons and the examination of two visualization methods(probability density function(PDF)and principal component analysis(PCA)),it is demonstrated that the generated data is hard to distinguish from the real data.The application of generated data for training a battery state model using transfer learning is further evaluated.Specifically,Bi-GRU-based and Transformer-based methods are implemented on 2 separate datasets for estimating state of health(SOH)and state of charge(SOC),respectively.The results indicate that the proposed framework demonstrates satisfactory performance in different scenarios:for the data replacement scenario,where real data are removed and replaced with generated data,the state estimator accuracy decreases only slightly;for the data enhancement scenario,the estimator accuracy is further improved.The estimation accuracy of SOH and SOC is as low as 0.69%and 0.58%root mean square error(RMSE)after applying the proposed framework.This framework provides a reliable method for enriching battery measurement data.It is a generalized framework capable of generating a variety of time series data.展开更多
Large-scale point cloud datasets form the basis for training various deep learning networks and achieving high-quality network processing tasks.Due to the diversity and robustness constraints of the data,data augmenta...Large-scale point cloud datasets form the basis for training various deep learning networks and achieving high-quality network processing tasks.Due to the diversity and robustness constraints of the data,data augmentation(DA)methods are utilised to expand dataset diversity and scale.However,due to the complex and distinct characteristics of LiDAR point cloud data from different platforms(such as missile-borne and vehicular LiDAR data),directly applying traditional 2D visual domain DA methods to 3D data can lead to networks trained using this approach not robustly achieving the corresponding tasks.To address this issue,the present study explores DA for missile-borne LiDAR point cloud using a Monte Carlo(MC)simulation method that closely resembles practical application.Firstly,the model of multi-sensor imaging system is established,taking into account the joint errors arising from the platform itself and the relative motion during the imaging process.A distortion simulation method based on MC simulation for augmenting missile-borne LiDAR point cloud data is proposed,underpinned by an analysis of combined errors between different modal sensors,achieving high-quality augmentation of point cloud data.The effectiveness of the proposed method in addressing imaging system errors and distortion simulation is validated using the imaging scene dataset constructed in this paper.Comparative experiments between the proposed point cloud DA algorithm and the current state-of-the-art algorithms in point cloud detection and single object tracking tasks demonstrate that the proposed method can improve the network performance obtained from unaugmented datasets by over 17.3%and 17.9%,surpassing SOTA performance of current point cloud DA algorithms.展开更多
Rainfall-induced shallow landslides pose one of significant geological hazards,necessitating precise monitoring and prediction for effective disaster mitigation.Most studies on landslide prediction have focused on opt...Rainfall-induced shallow landslides pose one of significant geological hazards,necessitating precise monitoring and prediction for effective disaster mitigation.Most studies on landslide prediction have focused on optimizing machine learning(ML)algorithms,very limited attention has been paid to enhancing data quality for improved predictive performance.This study employs strategic data augmentation(DA)techniques to enhance the accuracy of shallow landslide prediction.Using five DA methods including singular spectrum analysis(SSA),moving averages(MA),wavelet denoising(WD),variational mode decomposition(VMD),and linear interpolation(LI),we utilize strategies such as smoothing,denoising,trend decomposition,and synthetic data generation to improve the training dataset.Four machine learning algorithms,i.e.artificial neural network(ANN),recurrent neural network(RNN),one-dimensional convolutional neural network(CNN1D),and long short-term memory(LSTM),are used to forecast landslide displacement.The case study of a landslide in southwest China shows the effectiveness of our approach in predicting landslide displacements,despite the inherent limitations of the monitoring dataset.VMD proves the most effective for smoothing and denoising,improving R^(2),RMSE,and MAPE by 172.16%,71.82%,and 98.9%,respectively.SSA addresses missing data,while LI is effective with limited data samples,improving metrics by 21.6%,52.59%,and 47.87%,respectively.This study demonstrates the potential of DA techniques to mitigate the impact of data defects on landslide prediction accuracy,with implications for similar cases.展开更多
The widespread adoption of tunnel boring machines(TBMs)has led to an increased focus on disc cutter wear,including both normal and abnormal types,for efficient and safe TBM excavation.However,abnormal wear has yet to ...The widespread adoption of tunnel boring machines(TBMs)has led to an increased focus on disc cutter wear,including both normal and abnormal types,for efficient and safe TBM excavation.However,abnormal wear has yet to be thoroughly investigated,primarily due to the complexity of considering mixed ground conditions and the imbalance in the number of instances between the two types of wear.This study developed a prediction model for abnormal TBM disc cutter wear,considering mixed ground conditions,by employing interpretable machine learning with data augmentation.An equivalent elastic modulus was used to consider the characteristics of mixed ground conditions,and wear data was obtained from 65 cutterhead intervention(CHI)reports covering both mixed ground and hard rock sections.With a balanced training dataset obtained by data augmentation,an extreme gradient boosting(XGB)model delivered acceptable results with an accuracy of 0.94,an F1-score of 0.808,and a recall of 0.8.In addition,the accuracy for each individual disc cutter exhibited low variability.When employing data augmentation,a significant improvement in recall was observed compared to when it was not used,although the difference in accuracy and F1-score was marginal.The subsequent model interpretation revealed the chamber pressure,cutter installation radius,and torque as significant contributors.Specifically,a threshold in chamber pressure was observed,which could induce abnormal wear.The study also explored how elevated values of these influential contributors correlate with abnormal wear.The proposed model offers a valuable tool for planning the replacement of abnormally worn disc cutters,enhancing the safety and efficiency of TBM operations.展开更多
With the emergence of new attack techniques,traffic classifiers usually fail to maintain the expected performance in real-world network environments.In order to have sufficient generalizability to deal with unknown ma...With the emergence of new attack techniques,traffic classifiers usually fail to maintain the expected performance in real-world network environments.In order to have sufficient generalizability to deal with unknown malicious samples,they require a large number of new samples for retraining.Considering the cost of data collection and labeling,data augmentation is an ideal solution.We propose an optimized noise-based traffic data augmentation system,ONTDAS.The system uses a gradient-based searching algorithm and an improved Bayesian optimizer to obtain optimized noise.The noise is injected into the original samples for data augmentation.Then,an improved bagging algorithm is used to integrate all the base traffic classifiers trained on noised datasets.The experiments verify ONTDAS on 6 types of base classifiers and 4 publicly available datasets respectively.The results show that ONTDAS can effectively enhance the traffic classifiers’performance and significantly improve their generalizability on unknown malicious samples.The system can also alleviate dataset imbalance.Moreover,the performance of ONTDAS is significantly superior to the existing data augmentation methods mentioned.展开更多
Data augmentation plays an important role in training deep neural model by expanding the size and diversity of the dataset.Initially,data augmentation mainly involved some simple transformations of images.Later,in ord...Data augmentation plays an important role in training deep neural model by expanding the size and diversity of the dataset.Initially,data augmentation mainly involved some simple transformations of images.Later,in order to increase the diversity and complexity of data,more advanced methods appeared and evolved to sophisticated generative models.However,these methods required a mass of computation of training or searching.In this paper,a novel training-free method that utilises the Pre-Trained Segment Anything Model(SAM)model as a data augmentation tool(PTSAM-DA)is proposed to generate the augmented annotations for images.Without the need for training,it obtains prompt boxes from the original annotations and then feeds the boxes to the pre-trained SAM to generate diverse and improved annotations.In this way,annotations are augmented more ingenious than simple manipulations without incurring huge computation for training a data augmentation model.Multiple comparative experiments on three datasets are conducted,including an in-house dataset,ADE20K and COCO2017.On this in-house dataset,namely Agricultural Plot Segmentation Dataset,maximum improvements of 3.77%and 8.92%are gained in two mainstream metrics,mIoU and mAcc,respectively.Consequently,large vision models like SAM are proven to be promising not only in image segmentation but also in data augmentation.展开更多
Medical image classification is crucial in disease diagnosis,treatment planning,and clinical decisionmaking.We introduced a novel medical image classification approach that integrates Bayesian Random Semantic Data Aug...Medical image classification is crucial in disease diagnosis,treatment planning,and clinical decisionmaking.We introduced a novel medical image classification approach that integrates Bayesian Random Semantic Data Augmentation(BSDA)with a Vision Mamba-based model for medical image classification(MedMamba),enhanced by residual connection blocks,we named the model BSDA-Mamba.BSDA augments medical image data semantically,enhancing the model’s generalization ability and classification performance.MedMamba,a deep learning-based state space model,excels in capturing long-range dependencies in medical images.By incorporating residual connections,BSDA-Mamba further improves feature extraction capabilities.Through comprehensive experiments on eight medical image datasets,we demonstrate that BSDA-Mamba outperforms existing models in accuracy,area under the curve,and F1-score.Our results highlight BSDA-Mamba’s potential as a reliable tool for medical image analysis,particularly in handling diverse imaging modalities from X-rays to MRI.The open-sourcing of our model’s code and datasets,will facilitate the reproduction and extension of our work.展开更多
Mg alloys possess an inherent plastic anisotropy owing to the selective activation of deformation mechanisms depending on the loading condition.This characteristic results in a diverse range of flow curves that vary w...Mg alloys possess an inherent plastic anisotropy owing to the selective activation of deformation mechanisms depending on the loading condition.This characteristic results in a diverse range of flow curves that vary with a deformation condition.This study proposes a novel approach for accurately predicting an anisotropic deformation behavior of wrought Mg alloys using machine learning(ML)with data augmentation.The developed model combines four key strategies from data science:learning the entire flow curves,generative adversarial networks(GAN),algorithm-driven hyperparameter tuning,and gated recurrent unit(GRU)architecture.The proposed model,namely GAN-aided GRU,was extensively evaluated for various predictive scenarios,such as interpolation,extrapolation,and a limited dataset size.The model exhibited significant predictability and improved generalizability for estimating the anisotropic compressive behavior of ZK60 Mg alloys under 11 annealing conditions and for three loading directions.The GAN-aided GRU results were superior to those of previous ML models and constitutive equations.The superior performance was attributed to hyperparameter optimization,GAN-based data augmentation,and the inherent predictivity of the GRU for extrapolation.As a first attempt to employ ML techniques other than artificial neural networks,this study proposes a novel perspective on predicting the anisotropic deformation behaviors of wrought Mg alloys.展开更多
Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal depende...Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal dependence,and noise.Therefore,methodologies for data augmentation and conversion of time series data into images for analysis have been studied.This paper proposes a fault detection model that uses time series data augmentation and transformation to address the problems of data imbalance,temporal dependence,and robustness to noise.The method of data augmentation is set as the addition of noise.It involves adding Gaussian noise,with the noise level set to 0.002,to maximize the generalization performance of the model.In addition,we use the Markov Transition Field(MTF)method to effectively visualize the dynamic transitions of the data while converting the time series data into images.It enables the identification of patterns in time series data and assists in capturing the sequential dependencies of the data.For anomaly detection,the PatchCore model is applied to show excellent performance,and the detected anomaly areas are represented as heat maps.It allows for the detection of anomalies,and by applying an anomaly map to the original image,it is possible to capture the areas where anomalies occur.The performance evaluation shows that both F1-score and Accuracy are high when time series data is converted to images.Additionally,when processed as images rather than as time series data,there was a significant reduction in both the size of the data and the training time.The proposed method can provide an important springboard for research in the field of anomaly detection using time series data.Besides,it helps solve problems such as analyzing complex patterns in data lightweight.展开更多
Mechanically cleaved two-dimensional materials are random in size and thickness.Recognizing atomically thin flakes by human experts is inefficient and unsuitable for scalable production.Deep learning algorithms have b...Mechanically cleaved two-dimensional materials are random in size and thickness.Recognizing atomically thin flakes by human experts is inefficient and unsuitable for scalable production.Deep learning algorithms have been adopted as an alternative,nevertheless a major challenge is a lack of sufficient actual training images.Here we report the generation of synthetic two-dimensional materials images using StyleGAN3 to complement the dataset.DeepLabv3Plus network is trained with the synthetic images which reduces overfitting and improves recognition accuracy to over 90%.A semi-supervisory technique for labeling images is introduced to reduce manual efforts.The sharper edges recognized by this method facilitate material stacking with precise edge alignment,which benefits exploring novel properties of layered-material devices that crucially depend on the interlayer twist-angle.This feasible and efficient method allows for the rapid and high-quality manufacturing of atomically thin materials and devices.展开更多
Damage to parcels reduces customer satisfactionwith delivery services and increases return-logistics costs.This can be prevented by detecting and addressing the damage before the parcels reach the customer.Consequentl...Damage to parcels reduces customer satisfactionwith delivery services and increases return-logistics costs.This can be prevented by detecting and addressing the damage before the parcels reach the customer.Consequently,various studies have been conducted on deep learning techniques related to the detection of parcel damage.This study proposes a deep learning-based damage detectionmethod for various types of parcels.Themethod is intended to be part of a parcel information-recognition systemthat identifies the volume and shipping information of parcels,and determines whether they are damaged;this method is intended for use in the actual parcel-transportation process.For this purpose,1)the study acquired image data in an environment simulating the actual parcel-transportation process,and 2)the training dataset was expanded based on StyleGAN3 with adaptive discriminator augmentation.Additionally,3)a preliminary distinction was made between the appearance of parcels and their damage status to enhance the performance of the parcel damage detection model and analyze the causes of parcel damage.Finally,using the dataset constructed based on the proposed method,a damage type detection model was trained,and its mean average precision was confirmed.This model can improve customer satisfaction and reduce return costs for parcel delivery companies.展开更多
The study on ship wakes of synthetic aperture radar(SAR)images holds great importance in detecting ship targets in the ocean.In this study,we focus on the issues of low quantity and insufficient diversity in ship wake...The study on ship wakes of synthetic aperture radar(SAR)images holds great importance in detecting ship targets in the ocean.In this study,we focus on the issues of low quantity and insufficient diversity in ship wakes of SAR images,and propose a method of data augmentation of ship wakes in SAR images based on the improved cycle-consistent generative adversarial network(CycleGAN).The improvement measures mainly include two aspects:First,to enhance the quality of the generated images and guarantee a stable training process of the model,the least-squares loss is employed as the adversarial loss function;Second,the decoder of the generator is augmented with the convolutional block attention module(CBAM)to address the issue of missing details in the generated ship wakes of SAR images at the microscopic level.The experiment findings indicate that the improved CycleGAN model generates clearer ship wakes of SAR images,and outperforms the traditional CycleGAN models in both subjective and objective aspects.展开更多
Depth estimation is an important task in computer vision.Collecting data at scale for monocular depth estimation is challenging,as this task requires simultaneously capturing RGB images and depth information.Therefore...Depth estimation is an important task in computer vision.Collecting data at scale for monocular depth estimation is challenging,as this task requires simultaneously capturing RGB images and depth information.Therefore,data augmentation is crucial for this task.Existing data augmentationmethods often employ pixel-wise transformations,whichmay inadvertently disrupt edge features.In this paper,we propose a data augmentationmethod formonocular depth estimation,which we refer to as the Perpendicular-Cutdepth method.This method involves cutting realworld depth maps along perpendicular directions and pasting them onto input images,thereby diversifying the data without compromising edge features.To validate the effectiveness of the algorithm,we compared it with existing convolutional neural network(CNN)against the current mainstream data augmentation algorithms.Additionally,to verify the algorithm’s applicability to Transformer networks,we designed an encoder-decoder network structure based on Transformer to assess the generalization of our proposed algorithm.Experimental results demonstrate that,in the field of monocular depth estimation,our proposed method,Perpendicular-Cutdepth,outperforms traditional data augmentationmethods.On the indoor dataset NYU,our method increases accuracy from0.900 to 0.907 and reduces the error rate from0.357 to 0.351.On the outdoor dataset KITTI,our method improves accuracy from 0.9638 to 0.9642 and decreases the error rate from 0.060 to 0.0598.展开更多
The performance of deep learning models is heavily reliant on the quality and quantity of train-ing data.Insufficient training data will lead to overfitting.However,in the task of alert-situation text classification,i...The performance of deep learning models is heavily reliant on the quality and quantity of train-ing data.Insufficient training data will lead to overfitting.However,in the task of alert-situation text classification,it is usually difficult to obtain a large amount of training data.This paper proposes a text data augmentation method based on masked language model(MLM),aiming to enhance the generalization capability of deep learning models by expanding the training data.The method em-ploys a Mask strategy to randomly conceal words in the text,effectively leveraging contextual infor-mation to predict and replace masked words based on MLM,thereby generating new training data.Three Mask strategies of character level,word level and N-gram are designed,and the performance of each Mask strategy under different Mask ratios is analyzed and studied.The experimental results show that the performance of the word-level Mask strategy is better than the traditional data augmen-tation method.展开更多
The limited amount of data in the healthcare domain and the necessity of training samples for increased performance of deep learning models is a recurrent challenge,especially in medical imaging.Newborn Solutions aims...The limited amount of data in the healthcare domain and the necessity of training samples for increased performance of deep learning models is a recurrent challenge,especially in medical imaging.Newborn Solutions aims to enhance its non-invasive white blood cell counting device,Neosonics,by creating synthetic in vitro ultrasound images to facilitate a more efficient image generation process.This study addresses the data scarcity issue by designing and evaluating a continuous scalar conditional Generative Adversarial Network(GAN)to augment in vitro peritoneal dialysis ultrasound images,increasing both the volume and variability of training samples.The developed GAN architecture incorporates novel design features:varying kernel sizes in the generator’s transposed convolutional layers and a latent intermediate space,projecting noise and condition values for enhanced image resolution and specificity.The experimental results show that the GAN successfully generated diverse images of high visual quality,closely resembling real ultrasound samples.While visual results were promising,the use of GAN-based data augmentation did not consistently improve the performance of an image regressor in distinguishing features specific to varied white blood cell concentrations.Ultimately,while this continuous scalar conditional GAN model made strides in generating realistic images,further work is needed to achieve consistent gains in regression tasks,aiming for robust model generalization.展开更多
This paper investigates the problem of data scarcity in spectrum prediction.A cognitive radio equipment may frequently switch the target frequency as the electromagnetic environment changes.The previously trained mode...This paper investigates the problem of data scarcity in spectrum prediction.A cognitive radio equipment may frequently switch the target frequency as the electromagnetic environment changes.The previously trained model for prediction often cannot maintain a good performance when facing small amount of historical data of the new target frequency.Moreover,the cognitive radio equipment usually implements the dynamic spectrum access in real time which means the time to recollect the data of the new task frequency band and retrain the model is very limited.To address the above issues,we develop a crossband data augmentation framework for spectrum prediction by leveraging the recent advances of generative adversarial network(GAN)and deep transfer learning.Firstly,through the similarity measurement,we pre-train a GAN model using the historical data of the frequency band that is the most similar to the target frequency band.Then,through the data augmentation by feeding the small amount of the target data into the pre-trained GAN,temporal-spectral residual network is further trained using deep transfer learning and the generated data with high similarity from GAN.Finally,experiment results demonstrate the effectiveness of the proposed framework.展开更多
基金supported by the Jiangsu Province IUR Cooperation Project (No.BY2021258)the Wuxi Science and Technology Development Fund Project (No.G20212028)。
文摘An improved cycle-consistent generative adversarial network(CycleGAN) method for defect data augmentation based on feature fusion and self attention residual module is proposed to address the insufficiency of defect sample data for light guide plate(LGP) in production,as well as the problem of minor defects.Two optimizations are made to the generator of CycleGAN:fusion of low resolution features obtained from partial up-sampling and down-sampling with high-resolution features,combination of self attention mechanism with residual network structure to replace the original residual module.Qualitative and quantitative experiments were conducted to compare different data augmentation methods,and the results show that the defect images of the LGP generated by the improved network were more realistic,and the accuracy of the you only look once version 5(YOLOv5) detection network for the LGP was improved by 5.6%,proving the effectiveness and accuracy of the proposed method.
文摘The integration of image analysis through deep learning(DL)into rock classification represents a significant leap forward in geological research.While traditional methods remain invaluable for their expertise and historical context,DL offers a powerful complement by enhancing the speed,objectivity,and precision of the classification process.This research explores the significance of image data augmentation techniques in optimizing the performance of convolutional neural networks(CNNs)for geological image analysis,particularly in the classification of igneous,metamorphic,and sedimentary rock types from rock thin section(RTS)images.This study primarily focuses on classic image augmentation techniques and evaluates their impact on model accuracy and precision.Results demonstrate that augmentation techniques like Equalize significantly enhance the model's classification capabilities,achieving an F1-Score of 0.9869 for igneous rocks,0.9884 for metamorphic rocks,and 0.9929 for sedimentary rocks,representing improvements compared to the baseline original results.Moreover,the weighted average F1-Score across all classes and techniques is 0.9886,indicating an enhancement.Conversely,methods like Distort lead to decreased accuracy and F1-Score,with an F1-Score of 0.949 for igneous rocks,0.954 for metamorphic rocks,and 0.9416 for sedimentary rocks,exacerbating the performance compared to the baseline.The study underscores the practicality of image data augmentation in geological image classification and advocates for the adoption of DL methods in this domain for automation and improved results.The findings of this study can benefit various fields,including remote sensing,mineral exploration,and environmental monitoring,by enhancing the accuracy of geological image analysis both for scientific research and industrial applications.
基金supported by National Natural Science Foundation of China(61673186 and 61871196)Beijing Normal University Education Reform Project(jx2024040)Guangdong Undergraduate Universities Teaching Quality and Reform Project(jx2024309).
文摘Data augmentation plays an important role in boosting the performance of 3D models,while very few studies handle the 3D point cloud data with this technique.Global augmentation and cut-paste are commonly used augmentation techniques for point clouds,where global augmentation is applied to the entire point cloud of the scene,and cut-paste samples objects from other frames into the current frame.Both types of data augmentation can improve performance,but the cut-paste technique cannot effectively deal with the occlusion relationship between the foreground object and the background scene and the rationality of object sampling,which may be counterproductive and may hurt the overall performance.In addition,LiDAR is susceptible to signal loss,external occlusion,extreme weather and other factors,which can easily cause object shape changes,while global augmentation and cut-paste cannot effectively enhance the robustness of the model.To this end,we propose Syn-Aug,a synchronous data augmentation framework for LiDAR-based 3D object detection.Specifically,we first propose a novel rendering-based object augmentation technique(Ren-Aug)to enrich training data while enhancing scene realism.Second,we propose a local augmentation technique(Local-Aug)to generate local noise by rotating and scaling objects in the scene while avoiding collisions,which can improve generalisation performance.Finally,we make full use of the structural information of 3D labels to make the model more robust by randomly changing the geometry of objects in the training frames.We verify the proposed framework with four different types of 3D object detectors.Experimental results show that our proposed Syn-Aug significantly improves the performance of various 3D object detectors in the KITTI and nuScenes datasets,proving the effectiveness and generality of Syn-Aug.On KITTI,four different types of baseline models using Syn-Aug improved mAP by 0.89%,1.35%,1.61%and 1.14%respectively.On nuScenes,four different types of baseline models using Syn-Aug improved mAP by 14.93%,10.42%,8.47%and 6.81%respectively.The code is available at https://github.com/liuhuaijjin/Syn-Aug.
文摘Bird species classification is not only a challenging topic in artificial intelligence but also a domain closely related to environmental protection and ecological research.Additionally,performing edge computing on low-level devices using small neural networks can be an important research direction.In this paper,we use the EfficientNetV2B0 model for bird species classification,applying transfer learning on a dataset of 525 bird species.We also employ the BiRefNet model to remove backgrounds from images in the training set.The generated background-removed images are mixed with the original training set as a form of data augmentation.We aim for these background-removed images to help the model focus on key features,and by combining data augmentation with transfer learning,we trained a highly accurate and efficient bird species classification model.The training process is divided into a transfer learning stage and a fine-tuning stage.In the transfer learning stage,only the newly added custom layers are trained;while in the fine-tuning stage,all pre-trained layers except for the batch normalization layers are fine-tuned.According to the experimental results,the proposed model not only has an advantage in size compared to other models but also outperforms them in various metrics.The training results show that the proposed model achieved an accuracy of 99.54%and a precision of 99.62%,demonstrating that it achieves both lightweight design and high accuracy.To confirm the credibility of the results,we use heatmaps to interpret the model.The heatmaps show that our model can clearly highlight the image feature area.In addition,we also perform the 10-fold cross-validation on the model to verify its credibility.Finally,this paper proposes a model with low training cost and high accuracy,making it suitable for deployment on edge computing devices to provide lighter and more convenient services.
基金funded by the Bavarian State Ministry of Science,Research and Art(Grant number:H.2-F1116.WE/52/2)。
文摘In order to address the widespread data shortage problem in battery research,this paper proposes a generative adversarial network model that combines it with deep convolutional networks,the Wasserstein distance,and the gradient penalty to achieve data augmentation.To lower the threshold for implementing the proposed method,transfer learning is further introduced.The W-DC-GAN-GP-TL framework is thereby formed.This framework is evaluated on 3 different publicly available datasets to judge the quality of generated data.Through visual comparisons and the examination of two visualization methods(probability density function(PDF)and principal component analysis(PCA)),it is demonstrated that the generated data is hard to distinguish from the real data.The application of generated data for training a battery state model using transfer learning is further evaluated.Specifically,Bi-GRU-based and Transformer-based methods are implemented on 2 separate datasets for estimating state of health(SOH)and state of charge(SOC),respectively.The results indicate that the proposed framework demonstrates satisfactory performance in different scenarios:for the data replacement scenario,where real data are removed and replaced with generated data,the state estimator accuracy decreases only slightly;for the data enhancement scenario,the estimator accuracy is further improved.The estimation accuracy of SOH and SOC is as low as 0.69%and 0.58%root mean square error(RMSE)after applying the proposed framework.This framework provides a reliable method for enriching battery measurement data.It is a generalized framework capable of generating a variety of time series data.
基金Postgraduate Innovation Top notch Talent Training Project of Hunan Province,Grant/Award Number:CX20220045Scientific Research Project of National University of Defense Technology,Grant/Award Number:22-ZZCX-07+2 种基金New Era Education Quality Project of Anhui Province,Grant/Award Number:2023cxcysj194National Natural Science Foundation of China,Grant/Award Numbers:62201597,62205372,1210456foundation of Hefei Comprehensive National Science Center,Grant/Award Number:KY23C502。
文摘Large-scale point cloud datasets form the basis for training various deep learning networks and achieving high-quality network processing tasks.Due to the diversity and robustness constraints of the data,data augmentation(DA)methods are utilised to expand dataset diversity and scale.However,due to the complex and distinct characteristics of LiDAR point cloud data from different platforms(such as missile-borne and vehicular LiDAR data),directly applying traditional 2D visual domain DA methods to 3D data can lead to networks trained using this approach not robustly achieving the corresponding tasks.To address this issue,the present study explores DA for missile-borne LiDAR point cloud using a Monte Carlo(MC)simulation method that closely resembles practical application.Firstly,the model of multi-sensor imaging system is established,taking into account the joint errors arising from the platform itself and the relative motion during the imaging process.A distortion simulation method based on MC simulation for augmenting missile-borne LiDAR point cloud data is proposed,underpinned by an analysis of combined errors between different modal sensors,achieving high-quality augmentation of point cloud data.The effectiveness of the proposed method in addressing imaging system errors and distortion simulation is validated using the imaging scene dataset constructed in this paper.Comparative experiments between the proposed point cloud DA algorithm and the current state-of-the-art algorithms in point cloud detection and single object tracking tasks demonstrate that the proposed method can improve the network performance obtained from unaugmented datasets by over 17.3%and 17.9%,surpassing SOTA performance of current point cloud DA algorithms.
基金supported by the National Natural Science Foundation of China(Grant No.42101089)Sichuan Science and Technology Program(2022YFS0586)the Open Fund of Key Laboratory of Mountain Hazards and Earth Surface Processes Chinese Academy of Sciences.
文摘Rainfall-induced shallow landslides pose one of significant geological hazards,necessitating precise monitoring and prediction for effective disaster mitigation.Most studies on landslide prediction have focused on optimizing machine learning(ML)algorithms,very limited attention has been paid to enhancing data quality for improved predictive performance.This study employs strategic data augmentation(DA)techniques to enhance the accuracy of shallow landslide prediction.Using five DA methods including singular spectrum analysis(SSA),moving averages(MA),wavelet denoising(WD),variational mode decomposition(VMD),and linear interpolation(LI),we utilize strategies such as smoothing,denoising,trend decomposition,and synthetic data generation to improve the training dataset.Four machine learning algorithms,i.e.artificial neural network(ANN),recurrent neural network(RNN),one-dimensional convolutional neural network(CNN1D),and long short-term memory(LSTM),are used to forecast landslide displacement.The case study of a landslide in southwest China shows the effectiveness of our approach in predicting landslide displacements,despite the inherent limitations of the monitoring dataset.VMD proves the most effective for smoothing and denoising,improving R^(2),RMSE,and MAPE by 172.16%,71.82%,and 98.9%,respectively.SSA addresses missing data,while LI is effective with limited data samples,improving metrics by 21.6%,52.59%,and 47.87%,respectively.This study demonstrates the potential of DA techniques to mitigate the impact of data defects on landslide prediction accuracy,with implications for similar cases.
基金support of the“National R&D Project for Smart Construction Technology (Grant No.RS-2020-KA157074)”funded by the Korea Agency for Infrastructure Technology Advancement under the Ministry of Land,Infrastructure and Transport,and managed by the Korea Expressway Corporation.
文摘The widespread adoption of tunnel boring machines(TBMs)has led to an increased focus on disc cutter wear,including both normal and abnormal types,for efficient and safe TBM excavation.However,abnormal wear has yet to be thoroughly investigated,primarily due to the complexity of considering mixed ground conditions and the imbalance in the number of instances between the two types of wear.This study developed a prediction model for abnormal TBM disc cutter wear,considering mixed ground conditions,by employing interpretable machine learning with data augmentation.An equivalent elastic modulus was used to consider the characteristics of mixed ground conditions,and wear data was obtained from 65 cutterhead intervention(CHI)reports covering both mixed ground and hard rock sections.With a balanced training dataset obtained by data augmentation,an extreme gradient boosting(XGB)model delivered acceptable results with an accuracy of 0.94,an F1-score of 0.808,and a recall of 0.8.In addition,the accuracy for each individual disc cutter exhibited low variability.When employing data augmentation,a significant improvement in recall was observed compared to when it was not used,although the difference in accuracy and F1-score was marginal.The subsequent model interpretation revealed the chamber pressure,cutter installation radius,and torque as significant contributors.Specifically,a threshold in chamber pressure was observed,which could induce abnormal wear.The study also explored how elevated values of these influential contributors correlate with abnormal wear.The proposed model offers a valuable tool for planning the replacement of abnormally worn disc cutters,enhancing the safety and efficiency of TBM operations.
基金supported in part by the National Key Research and Development Program of China(No.2022YFB4500800)the National Science Foundation of China(No.42071431).
文摘With the emergence of new attack techniques,traffic classifiers usually fail to maintain the expected performance in real-world network environments.In order to have sufficient generalizability to deal with unknown malicious samples,they require a large number of new samples for retraining.Considering the cost of data collection and labeling,data augmentation is an ideal solution.We propose an optimized noise-based traffic data augmentation system,ONTDAS.The system uses a gradient-based searching algorithm and an improved Bayesian optimizer to obtain optimized noise.The noise is injected into the original samples for data augmentation.Then,an improved bagging algorithm is used to integrate all the base traffic classifiers trained on noised datasets.The experiments verify ONTDAS on 6 types of base classifiers and 4 publicly available datasets respectively.The results show that ONTDAS can effectively enhance the traffic classifiers’performance and significantly improve their generalizability on unknown malicious samples.The system can also alleviate dataset imbalance.Moreover,the performance of ONTDAS is significantly superior to the existing data augmentation methods mentioned.
基金Natural Science Foundation of Zhejiang Province,Grant/Award Number:LY23F020025Science and Technology Commissioner Program of Huzhou,Grant/Award Number:2023GZ42Sichuan Provincial Science and Technology Support Program,Grant/Award Numbers:2023ZHCG0005,2023ZHCG0008。
文摘Data augmentation plays an important role in training deep neural model by expanding the size and diversity of the dataset.Initially,data augmentation mainly involved some simple transformations of images.Later,in order to increase the diversity and complexity of data,more advanced methods appeared and evolved to sophisticated generative models.However,these methods required a mass of computation of training or searching.In this paper,a novel training-free method that utilises the Pre-Trained Segment Anything Model(SAM)model as a data augmentation tool(PTSAM-DA)is proposed to generate the augmented annotations for images.Without the need for training,it obtains prompt boxes from the original annotations and then feeds the boxes to the pre-trained SAM to generate diverse and improved annotations.In this way,annotations are augmented more ingenious than simple manipulations without incurring huge computation for training a data augmentation model.Multiple comparative experiments on three datasets are conducted,including an in-house dataset,ADE20K and COCO2017.On this in-house dataset,namely Agricultural Plot Segmentation Dataset,maximum improvements of 3.77%and 8.92%are gained in two mainstream metrics,mIoU and mAcc,respectively.Consequently,large vision models like SAM are proven to be promising not only in image segmentation but also in data augmentation.
文摘Medical image classification is crucial in disease diagnosis,treatment planning,and clinical decisionmaking.We introduced a novel medical image classification approach that integrates Bayesian Random Semantic Data Augmentation(BSDA)with a Vision Mamba-based model for medical image classification(MedMamba),enhanced by residual connection blocks,we named the model BSDA-Mamba.BSDA augments medical image data semantically,enhancing the model’s generalization ability and classification performance.MedMamba,a deep learning-based state space model,excels in capturing long-range dependencies in medical images.By incorporating residual connections,BSDA-Mamba further improves feature extraction capabilities.Through comprehensive experiments on eight medical image datasets,we demonstrate that BSDA-Mamba outperforms existing models in accuracy,area under the curve,and F1-score.Our results highlight BSDA-Mamba’s potential as a reliable tool for medical image analysis,particularly in handling diverse imaging modalities from X-rays to MRI.The open-sourcing of our model’s code and datasets,will facilitate the reproduction and extension of our work.
基金Korea Institute of Energy Technology Evaluation and Planning(KETEP)grant funded by the Korea government(Grant No.20214000000140,Graduate School of Convergence for Clean Energy Integrated Power Generation)Korea Basic Science Institute(National Research Facilities and Equipment Center)grant funded by the Ministry of Education(2021R1A6C101A449)the National Research Foundation of Korea grant funded by the Ministry of Science and ICT(2021R1A2C1095139),Republic of Korea。
文摘Mg alloys possess an inherent plastic anisotropy owing to the selective activation of deformation mechanisms depending on the loading condition.This characteristic results in a diverse range of flow curves that vary with a deformation condition.This study proposes a novel approach for accurately predicting an anisotropic deformation behavior of wrought Mg alloys using machine learning(ML)with data augmentation.The developed model combines four key strategies from data science:learning the entire flow curves,generative adversarial networks(GAN),algorithm-driven hyperparameter tuning,and gated recurrent unit(GRU)architecture.The proposed model,namely GAN-aided GRU,was extensively evaluated for various predictive scenarios,such as interpolation,extrapolation,and a limited dataset size.The model exhibited significant predictability and improved generalizability for estimating the anisotropic compressive behavior of ZK60 Mg alloys under 11 annealing conditions and for three loading directions.The GAN-aided GRU results were superior to those of previous ML models and constitutive equations.The superior performance was attributed to hyperparameter optimization,GAN-based data augmentation,and the inherent predictivity of the GRU for extrapolation.As a first attempt to employ ML techniques other than artificial neural networks,this study proposes a novel perspective on predicting the anisotropic deformation behaviors of wrought Mg alloys.
基金This research was financially supported by the Ministry of Trade,Industry,and Energy(MOTIE),Korea,under the“Project for Research and Development with Middle Markets Enterprises and DNA(Data,Network,AI)Universities”(AI-based Safety Assessment and Management System for Concrete Structures)(ReferenceNumber P0024559)supervised by theKorea Institute for Advancement of Technology(KIAT).
文摘Time-series data provide important information in many fields,and their processing and analysis have been the focus of much research.However,detecting anomalies is very difficult due to data imbalance,temporal dependence,and noise.Therefore,methodologies for data augmentation and conversion of time series data into images for analysis have been studied.This paper proposes a fault detection model that uses time series data augmentation and transformation to address the problems of data imbalance,temporal dependence,and robustness to noise.The method of data augmentation is set as the addition of noise.It involves adding Gaussian noise,with the noise level set to 0.002,to maximize the generalization performance of the model.In addition,we use the Markov Transition Field(MTF)method to effectively visualize the dynamic transitions of the data while converting the time series data into images.It enables the identification of patterns in time series data and assists in capturing the sequential dependencies of the data.For anomaly detection,the PatchCore model is applied to show excellent performance,and the detected anomaly areas are represented as heat maps.It allows for the detection of anomalies,and by applying an anomaly map to the original image,it is possible to capture the areas where anomalies occur.The performance evaluation shows that both F1-score and Accuracy are high when time series data is converted to images.Additionally,when processed as images rather than as time series data,there was a significant reduction in both the size of the data and the training time.The proposed method can provide an important springboard for research in the field of anomaly detection using time series data.Besides,it helps solve problems such as analyzing complex patterns in data lightweight.
基金Project supported by the National Key Research and Development Program of China(Grant No.2022YFB2803900)the National Natural Science Foundation of China(Grant Nos.61974075 and 61704121)+2 种基金the Natural Science Foundation of Tianjin Municipality(Grant Nos.22JCZDJC00460 and 19JCQNJC00700)Tianjin Municipal Education Commission(Grant No.2019KJ028)Fundamental Research Funds for the Central Universities(Grant No.22JCZDJC00460).
文摘Mechanically cleaved two-dimensional materials are random in size and thickness.Recognizing atomically thin flakes by human experts is inefficient and unsuitable for scalable production.Deep learning algorithms have been adopted as an alternative,nevertheless a major challenge is a lack of sufficient actual training images.Here we report the generation of synthetic two-dimensional materials images using StyleGAN3 to complement the dataset.DeepLabv3Plus network is trained with the synthetic images which reduces overfitting and improves recognition accuracy to over 90%.A semi-supervisory technique for labeling images is introduced to reduce manual efforts.The sharper edges recognized by this method facilitate material stacking with precise edge alignment,which benefits exploring novel properties of layered-material devices that crucially depend on the interlayer twist-angle.This feasible and efficient method allows for the rapid and high-quality manufacturing of atomically thin materials and devices.
基金supported by a Korea Agency for Infrastructure Technology Advancement(KAIA)grant funded by the Ministry of Land,Infrastructure,and Transport(Grant 1615013176)(https://www.kaia.re.kr/eng/main.do,accessed on 01/06/2024)supported by a Korea Evaluation Institute of Industrial Technology(KEIT)grant funded by the Korean Government(MOTIE)(141518499)(https://www.keit.re.kr/index.es?sid=a2,accessed on 01/06/2024).
文摘Damage to parcels reduces customer satisfactionwith delivery services and increases return-logistics costs.This can be prevented by detecting and addressing the damage before the parcels reach the customer.Consequently,various studies have been conducted on deep learning techniques related to the detection of parcel damage.This study proposes a deep learning-based damage detectionmethod for various types of parcels.Themethod is intended to be part of a parcel information-recognition systemthat identifies the volume and shipping information of parcels,and determines whether they are damaged;this method is intended for use in the actual parcel-transportation process.For this purpose,1)the study acquired image data in an environment simulating the actual parcel-transportation process,and 2)the training dataset was expanded based on StyleGAN3 with adaptive discriminator augmentation.Additionally,3)a preliminary distinction was made between the appearance of parcels and their damage status to enhance the performance of the parcel damage detection model and analyze the causes of parcel damage.Finally,using the dataset constructed based on the proposed method,a damage type detection model was trained,and its mean average precision was confirmed.This model can improve customer satisfaction and reduce return costs for parcel delivery companies.
基金Industry-University-Research Cooperation Fund Project of the Eighth Research Institute of China Aerospace Science and Technology Corporation(No.USCAST2021-5)Aeronautical Science Foundation of China(No.20220001057001)。
文摘The study on ship wakes of synthetic aperture radar(SAR)images holds great importance in detecting ship targets in the ocean.In this study,we focus on the issues of low quantity and insufficient diversity in ship wakes of SAR images,and propose a method of data augmentation of ship wakes in SAR images based on the improved cycle-consistent generative adversarial network(CycleGAN).The improvement measures mainly include two aspects:First,to enhance the quality of the generated images and guarantee a stable training process of the model,the least-squares loss is employed as the adversarial loss function;Second,the decoder of the generator is augmented with the convolutional block attention module(CBAM)to address the issue of missing details in the generated ship wakes of SAR images at the microscopic level.The experiment findings indicate that the improved CycleGAN model generates clearer ship wakes of SAR images,and outperforms the traditional CycleGAN models in both subjective and objective aspects.
基金the Grant of Program for Scientific ResearchInnovation Team in Colleges and Universities of Anhui Province(2022AH010095)The Grant ofScientific Research and Talent Development Foundation of the Hefei University(No.21-22RC15)+2 种基金The Key Research Plan of Anhui Province(No.2022k07020011)The Grant of Anhui Provincial940 CMC,2024,vol.79,no.1Natural Science Foundation,No.2308085MF213The Open Fund of Information Materials andIntelligent Sensing Laboratory of Anhui Province IMIS202205,as well as the AI General ComputingPlatform of Hefei University.
文摘Depth estimation is an important task in computer vision.Collecting data at scale for monocular depth estimation is challenging,as this task requires simultaneously capturing RGB images and depth information.Therefore,data augmentation is crucial for this task.Existing data augmentationmethods often employ pixel-wise transformations,whichmay inadvertently disrupt edge features.In this paper,we propose a data augmentationmethod formonocular depth estimation,which we refer to as the Perpendicular-Cutdepth method.This method involves cutting realworld depth maps along perpendicular directions and pasting them onto input images,thereby diversifying the data without compromising edge features.To validate the effectiveness of the algorithm,we compared it with existing convolutional neural network(CNN)against the current mainstream data augmentation algorithms.Additionally,to verify the algorithm’s applicability to Transformer networks,we designed an encoder-decoder network structure based on Transformer to assess the generalization of our proposed algorithm.Experimental results demonstrate that,in the field of monocular depth estimation,our proposed method,Perpendicular-Cutdepth,outperforms traditional data augmentationmethods.On the indoor dataset NYU,our method increases accuracy from0.900 to 0.907 and reduces the error rate from0.357 to 0.351.On the outdoor dataset KITTI,our method improves accuracy from 0.9638 to 0.9642 and decreases the error rate from 0.060 to 0.0598.
基金Supported by the Humanities and Social Sciences Research Project of the Ministry of Education(No.22YJA840004).
文摘The performance of deep learning models is heavily reliant on the quality and quantity of train-ing data.Insufficient training data will lead to overfitting.However,in the task of alert-situation text classification,it is usually difficult to obtain a large amount of training data.This paper proposes a text data augmentation method based on masked language model(MLM),aiming to enhance the generalization capability of deep learning models by expanding the training data.The method em-ploys a Mask strategy to randomly conceal words in the text,effectively leveraging contextual infor-mation to predict and replace masked words based on MLM,thereby generating new training data.Three Mask strategies of character level,word level and N-gram are designed,and the performance of each Mask strategy under different Mask ratios is analyzed and studied.The experimental results show that the performance of the word-level Mask strategy is better than the traditional data augmen-tation method.
文摘The limited amount of data in the healthcare domain and the necessity of training samples for increased performance of deep learning models is a recurrent challenge,especially in medical imaging.Newborn Solutions aims to enhance its non-invasive white blood cell counting device,Neosonics,by creating synthetic in vitro ultrasound images to facilitate a more efficient image generation process.This study addresses the data scarcity issue by designing and evaluating a continuous scalar conditional Generative Adversarial Network(GAN)to augment in vitro peritoneal dialysis ultrasound images,increasing both the volume and variability of training samples.The developed GAN architecture incorporates novel design features:varying kernel sizes in the generator’s transposed convolutional layers and a latent intermediate space,projecting noise and condition values for enhanced image resolution and specificity.The experimental results show that the GAN successfully generated diverse images of high visual quality,closely resembling real ultrasound samples.While visual results were promising,the use of GAN-based data augmentation did not consistently improve the performance of an image regressor in distinguishing features specific to varied white blood cell concentrations.Ultimately,while this continuous scalar conditional GAN model made strides in generating realistic images,further work is needed to achieve consistent gains in regression tasks,aiming for robust model generalization.
基金This work was supported by the Science and Technology Innovation 2030-Key Project of“New Generation Artificial Intelligence”of China under Grant 2018AAA0102303the Natural Science Foundation for Distinguished Young Scholars of Jiangsu Province(No.BK20190030)the National Natural Science Foundation of China(No.61631020,No.61871398,No.61931011 and No.U20B2038).
文摘This paper investigates the problem of data scarcity in spectrum prediction.A cognitive radio equipment may frequently switch the target frequency as the electromagnetic environment changes.The previously trained model for prediction often cannot maintain a good performance when facing small amount of historical data of the new target frequency.Moreover,the cognitive radio equipment usually implements the dynamic spectrum access in real time which means the time to recollect the data of the new task frequency band and retrain the model is very limited.To address the above issues,we develop a crossband data augmentation framework for spectrum prediction by leveraging the recent advances of generative adversarial network(GAN)and deep transfer learning.Firstly,through the similarity measurement,we pre-train a GAN model using the historical data of the frequency band that is the most similar to the target frequency band.Then,through the data augmentation by feeding the small amount of the target data into the pre-trained GAN,temporal-spectral residual network is further trained using deep transfer learning and the generated data with high similarity from GAN.Finally,experiment results demonstrate the effectiveness of the proposed framework.
