Wyniki wyszukiwania - Biblioteka Nauki

1

Techniki powiększania danych stosowane w uczeniu maszynowym

100%

Ścigaj J.

Analit

|

2024

|

tom Nr 14

126--130

PL

Celem niniejszego referatu jest przybliżenie technik powiększania danych stosowanych w uczeniu maszynowym. Dzięki generowaniu nowych danych, na podstawie istniejących można w sposób łatwy uzyskać nowe zdjęcia, teksty, a także dźwięki. W jaki sposób to się dzieje, a także jakie są dostępne techniki, opisano w poniższym artykule. Przedstawiono również wady i zalety technik.

EN

The aim of this paper is to present data augmentation techniques used in machine learning. By generating new data, you can easily obtain new photos, texts and sounds based on the existing ones. How this happens, as well as what techniques are available, are described in the article below. Advantages and disadvantages of the techniques are also presented.

2

Data augmentation techniques for transfer learning improvement in drill wear classification using convolutional neural network

94%

Kurek J. , Antoniuk I. , Górski J. , Jegorowa A. , Świderski B. , Kruk M. , Wieczorek G. , Pach J. , Orłowski A. , Aleksiejuk-Gawron J.

Machine Graphics and Vision

|

2019

|

tom Vol. 28, No. 1/4

3--12

EN

This paper presents an improved method for recognizing the drill state on the basisof hole images drilled in a laminated chipboard, using convolutional neural network (CNN) and data augmentation techniques. Three classes were used to describe the drill state: red - for drill that is worn out and should be replaced, yellow - for state in which the system should send a warning to the operator, indicating that this element should be checked manually, and green - denoting the drill that is still in good condition, which allows for further use in the production process. The presented method combines the advantages of transfer learning and data augmentation methods to improve the accuracy of the received evaluations. In contrast to the classical deep learning methods, transfer learning requires much smaller training data sets to achieve acceptable results. At the same time, data augmentation customized for drill wear recognition makes it possible to expand the original dataset and to improve the overall accuracy. The experiments performed have confirmed the suitability of the presented approach to accurate class recognition in the given problem, even while using a small original dataset.

3

Defect Detection Using Deep Learning-Based YOLOv3 in Cross-Sectional Image of Additive Manufacturing

83%

Choi B. , Choi Y. , Lee M.-G. , Kim J.-S. , Lee S.-W. , Jeon Y.

Archives of Metallurgy and Materials

|

2021

|

tom Vol. 66, iss. 4

1037--1041

EN

Deposition defects like porosity, crack and lack of fusion in additive manufacturing process is a major obstacle to commercialization of the process. Thus, metallurgical microscopy analysis has been mainly conducted to optimize process conditions by detecting and investigating the defects. However, these defect detection methods indicate a deviation from the operator’s experience. In this study, artificial intelligence based YOLOv3 of object detection algorithm was applied to avoid the human dependency. The algorithm aims to automatically find and label the defects. To enable the aim, 80 training images and 20 verification images were prepared, and they were amplified into 640 training images and 160 verification images using augmentation algorithm of rotation, movement and scale down, randomly. To evaluate the performance of the algorithm, total loss was derived as the sum of localization loss, confidence loss, and classification loss. In the training process, the total loss was 8.672 for the initial 100 sample images. However, the total loss was reduced to 5.841 after training with additional 800 images. For the verification of the proposed method, new defect images were input and then the mean Average Precision (mAP) in terms of precision and recall was 0.3795. Therefore, the detection performance with high accuracy can be applied to industry for avoiding human errors.

4

Towards textual data augmentation for neural networks: synonyms and maximum loss

83%

Jungiewicz M. , Smywiński-Pohl A.

Computer Science

|

2019

|

tom Vol. 20 (1)

57--83

EN

Data augmentation is one of the ways to deal with labeled data scarcity and overfitting. Both of these problems are crucial for modern deep-learning algorithms, which require massive amounts of data. The problem is better explored in the context of image analysis than for text; this work is a step forward to help close this gap. We propose a method for augmenting textual data when training convolutional neural networks for sentence classification. The augmentation is based on the substitution of words using a thesaurus as well as Princeton University's WordNet. Our method improves upon the baseline in most of the cases. In terms of accuracy, the best of the variants is 1.2% (pp.) better than the baseline.

5

Treating dataset imbalance in fetal echocardiography classification

83%

Gusmão G. F. , Oliveira R. C. d. , Barbosa C. R. H. , Raposo A. B.

Annals of Computer Science and Information Systems

|

2022

|

tom Vol. 32

3--9

EN

Deep learning has been a trending topic during the last few years, notably in medical imaging that employs neural networks for image manipulation, computer-aided detection of diseases, and many other tasks depending on the clinical practices. One possible application that would benefit from these methods is the fetal cardiac view classification, where these different views are useful to obtain valuable information about the patient's heart development. A trained network could help reduce variance in interpretation and speed up data annotation. Alas, in this context we can face two challenges: datasets may contain a lot of information not relevant to the outcome of the classifier's training, and the view classes may be unbalanced in the sense that certain classes may have much more samples than others. This paper presents a series of attempts to solve these issues and can be used as a practical guide for training viable classifiers in this context.

6

Classifiers ensemble of transfer learning for improved drill wear classification using convolutional neural network

83%

Kurek J. , Antoniuk I. , Górski J. , Jegorowa A. , Świderski B. , Kruk M. , Wieczorek G. , Pach J. , Orłowski A. , Aleksiejuk-Gawron J.

Machine Graphics and Vision

|

2019

|

tom Vol. 28, No. 1/4

13--23

EN

In this paper we introduce the enhanced drill wear recognition method, based on classifiers ensemble, obtained using transfer learning and data augmentation methods. Red, green and yellow classes are used to describe the current drill state. The first one corresponds to the case when drill should be immediately replaced. The second one denotes a tool that is still in a good condition. The final class refers to the case when a drill is suspected of being worn out, and a human expert evaluation would be required. The proposed algorithm uses three different, pretrained network models and adjusts them to the drill wear classification problem. To ensure satisfactory results, each of the methods used was required to achieve accuracy above 90% for the given classification task. Final evaluation is achieved by voting of all three classifiers. Since the initial data set was small (242 instances), the data augmentation method was used to artificially increase the total number of drill hole images. The experiments performed confirmed that the presented approach can achieve high accuracy, even with such a limited set of training data.

7

Mixup (sample pairing) can improve the performance of deep segmentation networks

83%

Isaksson L. J. , Summers P. , Raimondi S. , Gandini S. , Bhalerao A. , Marvaso G. , Petralia G. , Pepa M. , Jereczek-Fossa B. A.

Journal of Artificial Intelligence and Soft Computing Research

|

2022

|

tom Vol. 12, No. 1

29--39

EN

Researchers address the generalization problem of deep image processing networks mainly through extensive use of data augmentation techniques such as random flips, rotations, and deformations. A data augmentation technique called mixup, which constructs virtual training samples from convex combinations of inputs, was recently proposed for deep classification networks. The algorithm contributed to increased performance on classification in a variety of datasets, but so far has not been evaluated for image segmentation tasks. In this paper, we tested whether the mixup algorithm can improve the generalization performance of deep segmentation networks for medical image data. We trained a standard U-net architecture to segment the prostate in 100 T2-weighted 3D magnetic resonance images from prostate cancer patients, and compared the results with and without mixup in terms of Dice similarity coefficient and mean surface distance from a reference segmentation made by an experienced radiologist. Our results suggest that mixup offers a statistically significant boost in performance compared to non-mixup training, leading to up to 1.9% increase in Dice and a 10.9% decrease in surface distance. The mixup algorithm may thus offer an important aid for medical image segmentation applications, which are typically limited by severe data scarcity.

8

Cell image augmentation for classification task using GANs on Pap smear dataset

71%

Zak J. , Grzeszczyk M. K. , Pater A. , Roszkowiak L. , Siemion K. , Korzynska A.

Biocybernetics and Biomedical Engineering

|

2022

|

tom Vol. 42, no. 3

995--1011

EN

One of the solutions to the problem of insufficiently large training datasets in image processing is data augmentation. This process artificially extends the size of training datasets to avoid overfitting. Generative Adversarial Networks yield that become increasingly difficult to differentiate from real images, until the differentiation is no longer possible. Thus, artificial images closely resembling original ones can be generated. Inclusion of artificial images contributes to improving the training process. Medical domain is one of the areas where data acquisition is burdened by many procedures, laws, and prohibitions. As a result the potential size of collected datasets is reduced. This article presents the results of training Convolutional Neural Networks on a artificially extended image datasets. The resulting classification accuracy on a cell classification task of models trained with images generated using the proposed method were increased by up to 12.9% in comparison to that of the model trained only with original dataset from the HErlev Pap smear dataset.

9

Parkinson’s disease classificationbased onstacked denoising autoencoder

71%

Sukanya P. , Srinivasa Rao B.

Computer Science

|

2023

|

tom T. 24 (4)

491--511

EN

One of the most common neurological conditions caused by gradual brain degen-eration is Parkinson’s disease (PD). Although this neurological condition hasno known treatment, early detection and therapy can help patients improvetheir quality of life. An essential patient’s health record is made of medicalimages used to control, manage, and treat diseases. However, in computer-based diagnostics, disease classification is a difficult task because of the timeconsumption and high rate of false positive marks. To overcome this problem, this paper introduces a stacked denoising autoencoder (SDA) for Parkinson’s disease classification. In preprocessing, noise is reduced and important information is retained, resulting in increased performance and data augmentationis performed to avoid overfitting issues and increase the size of the dataset.The main aim of this paper is to derive an optimal feature selection design foran effective Parkinson’s disease classification. Improved Pigeon-Inspired Optimization (IPIO) algorithm is introduced to enhance the performance of the classifier. Thus, the classification result improved by the optimal features and also increased the sensitivity, accuracy, and specificity in the medical image diagnosis. The proposed scheme is implemented in PYTHON and compared with traditional feature selection models and other classification approaches. The efficacy of the performances is evaluated using a Parkinson’s Progression Markers Initiative (PPMI) dataset. The integration of the stacked denoising autoencoder and Improved pigeon inspired optimization method produced the greatest results, with 99.17% accuracy, 98.74% sensitivity, and 98.96% specificity. Furthermore, our finding outperforms the most recent research in the field.

10

Data augmentation for haar cascade based automobile detection

71%

Pawełczyk M.

Archiwum Motoryzacji

|

2018

|

tom Vol. 82, nr 4

117--129

EN

The article describes recent object detection methods with their main advantages and drawbacks and shows results of application of machine learning Haar Cascade algorithm for automobile detection. The article underlines problems related to the feature dataset generation and presents an overview of current dataset augmentation methods such as image mirroring, cropping, rotating, shearing and color modification. New methods fot image dataset augmentation, such as utilization of CAD models and Deep Learning solutions, are also proposed. In order to ensure low cost, real time detection machine learning based Haar Cascade detector has been proposed and tested on a custom dataset specifically created for dataset augmentation methods evalutation. Article provides all input parameters for detector training process, along with a brief description of object detection metrics. Finally the article presents results of the baseline detector and augumented calssificator created based on vertical image mirroring technique, for different dataset configurations. Algorithms performance for real time detection on high resolution images was also evaluated.

11

Conditional generative adversarial network based data augmentation for fault diagnosis of diesel engines applied with infrared thermography and deep convolutional neural network

71%

Wang R. , Jia X. , Liu Z. , Dong E. , Li S. , Cheng Z.

Eksploatacja i Niezawodność

|

2024

|

tom Vol. 26, no. 1

art. no. 175291

EN

This paper tries to introduce a new intelligent method for the early fault diagnosis of diesel engines. Firstly, infrared thermography (IRT) is introduced into diesel engine condition monitoring, then infrared images of diesel engines in four health states, such as normal condition, single-cylinder misfire, multi-cylinder misfire and air filter blockage, are collected and the region of interest (ROI) of infrared images are extracted. Next, conditional generative adversarial network (CGAN) is deployed to perform data augmentation on infrared image datasets. Then, deep convolutional neural network (DCNN) and Softmax regression (SR) classifier are used for automatically extracting infrared image fault features and pattern recognition, respectively. Finally, a comparison with three deep learning (DL) models is performed. The validation results show that the data augmentation method proposed in the paper can significantly improve the early fault diagnosis accuracy, and DCNN has the best fault diagnosis effect andresistance to temperature fluctuation interference among the four DL models.

12

Generalizing clustering inferenceswith ml augmentation ofordinal survey data

71%

Kumar B. , Kumar R.

Computer Science

|

2024

|

tom T. 25 (1)

63--93

EN

In this paper, we attempt to generalize the ability to achieve quality inferences of survey data for a larger population through data augmentation andunification. Data augmentation techniques have proven effective in enhancingmodels’ performance by expanding the dataset’s size. We employ ML dataaugmentation, unification, and clustering techniques. First, we augment thelimitedsurvey data size using data augmentation technique(s). Second, wecarry out data unification, followed by clustering for inferencing. We took twobenchmark survey datasets to demonstrate the effectiveness of augmentationand unification. The first dataset contains information on aspiring studententrepreneurs’ characteristics, while the second dataset comprises survey datarelated to breast cancer. We compare the inferences drawn from the originalsurvey data with those derived from the transformed data using the proposedscheme. The results of this study indicate that the machine learning approach,data augmentation with the unification of data followed by clustering, can bebeneficial for generalizing the inferences drawn from the survey data.

13

Detection of Modic changes in MR images of spine using local binary patterns

59%

Athertya J. S. , Saravana Kumar G. , Govindaraj J.

Biocybernetics and Biomedical Engineering

|

2019

|

tom Vol. 39, no. 1

17--29

EN

Background and objective: With increase in prevalence of lower back pain, fast and reliable computer aided methods for clinical diagnosis associated with the same is needed for improving the healthcare reach. The magnetic resonance images exhibit a change in signal intensity on the vertebral body close to end plates, which are termed as Modic changes (MC), and are known to be clear indicators of lower back pain. The current work deals with computer aided methods for automating the classification of signal changes between normal and degenerate cases so as to aid physicians in precise and suitable diagnosis for the ailment. Methods: In order to detect Modic changes in vertebrae, initially the vertebrae are segmented from sagittal MR T1 and T2 imaged using a semi automatic cellular automata based segmentation. This is followed by textural feature extraction using Local Binary Patterns (LBP) and its variants. Various classifiers based on machine learning approaches using Random Forest, kNN, Bayes and SVM were evaluated for its classification performance. Since medical image dataset in general have bias towards healthy and diseased state, data augmentation techniques were also employed. Results: The implemented method is tested and validated over a dataset containing 100 patients. The proposed framework achieves an accuracy of 81% and 91.7% with and without augmentation of data respectively. A comparative study with the state of art methods reported in literature shows that the method proposed in better in terms of computational cost without any compromise on classification accuracy. Conclusion: A novel approach to identify MC in vertebrae by exploiting textural features is proposed. This shall assist radiologists in detecting abnormalities and in treatment planning.

14

Classification of lung nodule malignancy in computed tomography imaging utilising generative adversarial networks and semi-supervised transfer learning

59%

Apostolopoulos I. D. , Papathanasiou N. D. , Panayiotakis G. S.

Biocybernetics and Biomedical Engineering

|

2021

|

tom Vol. 41, no. 4

1243--1257

EN

The pulmonary nodules’ malignancy rating is commonly confined in patient follow-up; examining the nodule’s activity is estimated with the Positron Emission Tomography (PET) system or biopsy. However, these strategies are usually after the initial detection of the malignant nodules acquired from the Computed Tomography (CT) scan. In this study, a Deep Learning methodology to address the challenge of the automatic characterisation of Solitary Pulmonary Nodules (SPN) detected in CT scans is proposed. The research methodology is based on Convolutional Neural Networks, which have proven to be excellent automatic feature extractors for medical images. The publicly available CT dataset, called Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI), and a small CT scan dataset derived from a PET/CT system, is considered the classification target. New, realistic nodule representations are generated employing Deep Convolutional Generative Adversarial Networks to circumvent the shortage of large-scale data to train robust CNNs. Besides, a hierarchical CNN called Feature Fusion VGG19 (FF-VGG19) was developed to enhance feature extraction of the CNN proposed by the Visual Geometry Group (VGG). Moreover, the generated nodule images are separated into two classes by utilising a semi-supervised approach, called self-training, to tackle weak labelling due to DC-GAN inefficiencies. The DC-GAN can generate realistic SPNs, as the experts could only distinguish 23% of the synthetic nodule images. As a result, the classification accuracy of FF-VGG19 on the LIDCIDRI dataset increases by +7%, reaching 92.07%, while the classification accuracy on the CT dataset is increased by 5%, reaching 84,3%.

15

Machine learning for two-phase gas-liquid flow regime evaluation based on raw 3D ECT measurement data

59%

Wajman R. , Nowakowski J. , Łukiański M. , Banasiak R.

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2024

|

tom Vol. 72, nr 1

art. no. e148842

EN

This paper presents a study on applying machine learning algorithms for the classification of a two-phase flow regime and its internal structures. This research results may be used in adjusting optimal control of air pressure and liquid flow rate to pipeline and process vessels. To achieve this goal the model of an artificial neural network was built and trained using measurement data acquired from a 3D electrical capacitance tomography (ECT) measurement system. Because the set of measurement data collected to build the AI model was insufficient, a novel approach dedicated to data augmentation had to be developed. The main goal of the research was to examine the high adaptability of the artificial neural network (ANN) model in the case of emergency state and measurement system errors. Another goal was to test if it could resist unforeseen problems and correctly predict the flow type or detect these failures. It may help to avoid any pernicious damage and finally to compare its accuracy to the fuzzy classifier based on reconstructed tomography images – authors’ previous work.

16

Variational mode decomposition-based seizure classification using Bayesian regularized shallow neural network

59%

Yadav V. P. , Sharma K. K.

Biocybernetics and Biomedical Engineering

|

2021

|

tom Vol. 41, no. 2

402--418

EN

This work presents a new epileptic seizures epoch classification scheme. Variational mode decomposition (VMD), has been explored for non-recursively decomposing the electroencephalogram (EEG) signals into fourteen band limited intrinsic mode functions (IMFs). Data augmentation (DA), has been used for handling unbalanced classification problem. Normalized energy, fractal dimension, number of peaks, and prominence parameters were computed from the band-limited IMFs for the discrimination of seizure and non-seizure epochs. Bayesian regularized shallow neural network (BR-SNNs) and six other well-known classifiers were tested. Sensitivity, specificity, and accuracy have been used as performance metrics. This study includes two different epoch lengths of 1-second and 2-seconds. A total of 32 test cases for both, class balanced and unbalanced classification problems have been taken for the performance evaluation. The best performance obtained is 100% for all the three metrics from the test cases of database-2 and 3. For database-1, average sensitivity, specificity, and accuracy of 99.71, 99.75, and 99.73% have been achieved, respectively for the 1-second epoch. The presented work shows better performance results compared to many previously reported works.

17

Mediastinal lymph node malignancy detection in computed tomography images using fully convolutional network

59%

Tekchandani H. , Verma S. , Londhe N. D.

Biocybernetics and Biomedical Engineering

|

2020

|

tom Vol. 40, no. 1

187--199

EN

Differential diagnosis of malignant and benign mediastinal lymph nodes (LNs) through invasive pathological tests is a complex and painful procedure because of sophisticated anatomical locations of LNs in the chest. The image based automatic machine learning techniques have been attempted in the past for malignancy detection. But these conventional methods suffer from complex selection of hand-crafted features and trade-off between performance parameters due to them. Today deep learning approaches are out-performing conventional machine learning techniques and able to overcome these issues. However, the existing convolutional neural network (CNN) based models also are prone to overfitting because of fully connected (FC) layers. Therefore, in this paper authors have proposed a fully convolutional network (FCN) based deep learning model for lymph nodes malignancy detection in computed tomography (CT) images. Moreover, the proposed FCN has been customized with batch normalization and advanced activation function Leaky ReLU to accelerate the training and to overcome the problem of dying ReLU, respectively. The performance of the proposed FCN has been also tuned to its best for smaller data size using data augmentation methods. The generalization of the proposed model is tested using the network parameter variation. To understand the reliability of the proposed model, it has also been compared with state-of-art related deep learning networks. The proposed FCN model has achieved an average accuracy, sensitivity, specificity, and area under curve as 90.28%, 90.63%, 89.95%, and 0.90, respectively. Our results also confirms the successful usabilility of augmentation methods for working on smaller datasets and deep learning approaches.

18

A deep Residual U-Net convolutional neural network for automated lung segmentation in computed tomography images

59%

Khanna A. , Londhe N. D. , Gupta S. , Semwal A.

Biocybernetics and Biomedical Engineering

|

2020

|

tom Vol. 40, no. 3

1314--1327

EN

To improve the early diagnosis and treatment of lung diseases automated lung segmentation from CT images is a crucial task for clinical decision. The segmentation of the lung region from the CT scans is a very challenging task due to the irregular shape and size of lungs, low contrast and fuzzy boundaries of the lung. The manual segmentation of lung CT images is a laborious task. Therefore, various approaches are suggested by the researcher in the recent past for the automated lung segmentation. However, the existing approaches either utilize low-level handcraft features or CNN based Fully Convolutional Networks. The low-level hand-craft feature-based approaches lead to poor generalization, while the shallower networks are unable to extract more discriminative features. Hence, in this study, we have implemented a deep learning-based architecture called Residual U-Net with a false-positive removal algorithm for lung CT segmentation. Here, we have suggested that learning from a substantially deeper network with residual units can extract more discriminative feature representation as compared to shallow network for lung segmentation. To take full advantage of the deeper network, we have utilized a set of schemes to ensure efficient training. First, we have implemented a U-Net architecture with residual block to overcome the problem of performance degradation. Further, various data augmentation techniques are utilized to improve the generalization capability of the proposed method. The experimental results show that the proposed method achieved competitive results over the existing methods with DSC of 98.63%, 99.62% and 98.68% for LUNA16, VESSEL12 and HUG-ILD dataset respectively.

19

BiLSTM with Data Augmentation using Interpolation Methods to Improve Early Detection of Parkinson Disease

48%

Abayomi-Alli O. O. , Damaševičius R. , Maskeliūnas R. , Abayomi-Alli A.

Annals of Computer Science and Information Systems

|

2020

|

tom Vol. 21

371--380

EN

The lack of dopamine in the human brain is the cause of Parkinson disease (PD) which is a degenerative disorder common globally to older citizens. However, late detection of this disease before the first clinical diagnosis has led to increased mortality rate. Research effort towards the early detection of PD has encountered challenges such as: small dataset size, class imbalance, overfitting, high false detection rate, model complexity, etc. This paper aims to improve early detection of PD using machine learning through data augmentation for very small datasets. We propose using Spline interpolation and Piecewise Cubic Hermite Interpolating Polynomial (Pchip) interpolation methods to generate synthetic data instances. We further investigate on reducing dimensionality of features for effective and real-time classification while considering computational complexity of implementation on real-life mobile phones. For classification we use Bidirectional LSTM (BiLSTM) deep learning network and compare the results with traditional machine learning algorithms like Support Vector Machine (SVM), Decision Tree, Logistic regression, KNN and Ensemble bagged tree. For experimental validation we use the Oxford Parkinson disease dataset with 195 data samples, which we have augmented with 571 synthetic data samples. The results for BiLSTM shows that even with a holdout of 90%, the model was still able to effectively recognize PD with an average accuracy for ten rounds experiment using 22 features as 82.86%, 97.1\%, and 96.37% for original, augmented (Spline) and augmented (Pchip) datasets, respectively. Our results show that proposed data augmentation schemes have significantly (p < 0.001) improved the accuracy of PD recognition on a small dataset using both classical machine learning models and BiLSTM.

20

Analysis of the Impact of Data Augmentation on the Performance of Deep Learning Models in Multispectral Food Authenticity Identification

48%

Zhang Y. , Yilmaz A. , Popa M. , Brewster C.

Annals of Computer Science and Information Systems

|

2023

|

tom Vol. 35

823--832

EN

Food authenticity is a significant concern in the meat industry, demanding effective detection methods. This study explores the use of multispectral imaging (MSI) and deep learning for meat adulteration detection. We evaluate different deep learning models using transfer learning and preprocessing techniques in a multi-level adulteration classification task. In addition, we propose a novel approach called one-band mixed augmentation for band selection in MSI data, which outperforms traditional reflectance-based feature selection and enhances model robustness. Furthermore, employing the nine-crop approach for dataset augmentation improved the accuracy from 0.63 to 0.74 for DenseNet201 model without transfer learning. This research contributes to advancing food safety assessment practices and provides insights into the application of deep learning for preventing food adulteration. The proposed one-band mixed augmentation approach offers a novel strategy for handling band selection challenges in MSI data analysis.