Wyniki wyszukiwania - BazTech

1

Deep learning method for classifying items into categories for dutch auctions

Bobulski Janusz, Szymoniak Sabina

Computer Assisted Methods in Engineering and Science

|

2024

|

Vol. 31, no. 1

67--79

EN

Artificial Intelligence (AI) methods are widely used in our lives (phones, social media, self-driving cars, and e-commerce). In AI methods, we can find convolutional neural networks (CNN). First of all, we can use these networks to analyze images. This paper presents a method for classifying items into particular categories on an auction site. The technique prompts the seller to which category assign the item when creating a new auction. We choose a neural network with a number of image convolution layers as the best available approach to address this task. All tests were carried out in the Matlab environment using GPU and CPU. Then, the tested and verified solution was implemented in the TensorFlow environment with a CPU processor. Thanks to the cross-validation method, the effectiveness of the recognition system was fully verified in several stages. We obtained promising results. Consequently, we implemented the developed method by adding a new sales offer on the Clemens website.

2

Convolutional neural networks in the SSI analysis for mine-induced vibrations

Zając Maciej, Kuźniar Krystyna

Computer Assisted Methods in Engineering and Science

|

2024

|

Vol. 31, no. 1

3--28

EN

Deep neural networks (DNNs) have recently become one of the most often used softcomputational tools for numerical analysis. The huge success of DNNs in the field of imageprocessing is associated with the use of convolutional neural networks (CNNs). CNNs,thanks to their characteristic structure, allow for the effective extraction of multi-layerfeatures. In this paper, the application of CNNs to one of the important soil-structureinteraction (SSI) problems, i.e., the analysis of vibrations transmission from the free-field next to a building to the building foundation, is presented in the case of mine-induced vibrations. To achieve this, the dataset from in-situ experimental measurements,containing 1D ground acceleration records, was converted into 2D spectrogram imagesusing either Fourier transform or continuous wavelet transform. Next, these images wereused as input for a pre-trained CNN. The output is a ratio of maximal vibration valuesrecorded simultaneously on the building foundation and on the ground. Therefore, the lastlayer of the CNN had to be changed from a classification to a regression one. The obtainedresults indicate the suitability of CNN for the analyzed problem.

3

FedAssess: analysis for efficient communication and security algorithms over various federated learning frameworks and mitigation of label-flipping attack

Anusuya R., Karthika Renuka Dhanaraj

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2024

|

Vol. 72, nr 3

art. no. e148944

EN

Federated learning is an upcoming concept used widely in distributed machine learning. Federated learning (FL) allows a large number of users to learn a single machine learning model together while the training data is stored on individual user devices. Nonetheless, federated learning lessens threats to data privacy. Based on iterative model averaging, our study suggests a feasible technique for the federated learning of deep networks with improved security and privacy. We also undertake a thorough empirical evaluation while taking various FL frameworks and averaging algorithms into consideration. Secure multi party computation, secure aggregation, and differential privacy are implemented to improve the security and privacy in a federated learning environment. In spite of advancements, concerns over privacy remain in FL, as the weights or parameters of a trained model may reveal private information about the data used for training. Our work demonstrates that FL can be prone to label-flipping attack and a novel method to prevent label-flipping attack has been proposed. We compare standard federated model aggregation and optimization methods, FedAvg and FedProx using benchmark data sets. Experiments are implemented in two different FL frameworks – Flower and PySyft and the results are analyzed. Our experiments confirm that classification accuracy increases in FL framework over a centralized model and the model performance is better after adding all the security and privacy algorithms. Our work has proved that deep learning models perform well in FL and also is secure.

4

An improved facial image retrieval using hybrid cheetah optimization algorithm

Balasubramanian C., Raja Sekar J.

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2024

|

Vol. 72, nr 2

art. no. e148942

EN

With the advent of social media, the volume of photographs uploaded on the internet has increased exponentially. The task of efficiently recognizing and retrieving human facial images is inevitable and essential at this time. In this work, a feature selection approach for recognizing and retrieving human face images using hybrid cheetah optimization algorithm is proposed. The deep feature extraction from the images is done using deep convolutional neural networks. Hybrid cheetah optimization algorithm, an improvised version of cheetah optimization algorithm fused with genetic algorithm is used, to choose optimum features from the extracted deep features. The chosen features are used for finding the best-matching images from the image database. The image matching is performed by approximate nearest neighbor search for the query image over the image database and similar images are retrieved. By constructing a k-NN graph for the images, the efficiency of image retrieval is enhanced. The proposed system performance is evaluated against benchmark datasets such as LFW, MultiePie, ColorFERET, DigiFace-1M and CelebA. The evaluation results show that the proposed methodology is superior to various existing methodologies.

5

Travel time picking of ambient noise cross-correlation using a deep neural network combining convolutional neural networks and Transformer

Jin Chunwei, Ye Fang, Zhang Han, Bao Xue

Acta Geophysica

|

2024

|

Vol. 72, no. 1

97--114

EN

The travel time of ambient noise cross-correlation is widely used in geophysics, but traditional methods for picking the travel time of correlation are either difficult to be applied to data with low signal-to-noise ratio (SNR), or make some assumptions which fail to be achieved in many realistic situations, or require a lot of complex calculations. Here, we present a neural network based on convolutional neural networks (CNN) and Transformer for the travel time picking of ambient noise crosscorrelation. CNNs expand the dimension of the vector of each time step for the input of Transformer. Transformer focuses the model’s attention on the key parts of the sequence. Model derives the travel time according to the attention. 102,000 cross-correlations are used to train the network. Compared with traditional methods, the approach is easy to use and has a better performance, especially for the low SNR data. Then, we test our model on another ambient noise cross-correlation dataset, which contains cross-correlations from different regions and at different scales. The model has good performance on the test dataset. It can be seen from the experiment that the travel time of the cross-correlation function of ambient noise with an average SNR as low as 9.3 can be picked. 97.2% of the picked travel times are accurate, and the positive and negative travel time of most cross-correlations are identical (90.2%). Our method can be applied to seismic instrument performance verification, seismic velocity imaging, source location and other applications for its good ability to pick travel time accurately.

6

A hybrid drug named entity recognition framework for real time pubmed data using deep learning and text summarization techniques

Mathur T., Raimond Kumudha, Deepakanmani S.

Przegląd Elektrotechniczny

|

2023

|

R. 99, nr 8

106--109

EN

Drug Named Entity Recognition (DNER) becomes indispensable for various medical relation extraction systems. Existing deep learning systems rely on the benchmark data for training as well as testing the model. However, it is very important to test on the real time data. In this research, we propose a hybrid DNER framework where we incorporate text summarization on real time data to create the test dataset. We have experimented with various text summarization techniques and found SciBERT model to give better results than other techniques.

PL

Rozpoznawanie jednostek o nazwie leku (DNER) staje się nieodzowny dla innych systemów ekstrakcji relacji medycznych. Istniejące systemy głębokiego uczenia się opierają się na danych porównawczych zarówno podczas szkolenia, jak i testowania modelu. Jednak bardzo ważne jest, aby testować dane w czasie rzeczywistym. W tym badaniu proponujemy hybrydową strukturę DNER, w której uwzględniamy podsumowanie tekstu na danych w czasie rzeczywistym w celu utworzenia zestawu danych testowych. Eksperymentowaliśmy z różnymi technikami podsumowania tekstu i stwierdziliśmy, że model BERT daje lepsze wyniki niż inne techniki.

7

Brain tumor classification in MRI imagesusing genetic algorithm appended CNN

Balamurugan Thiyagu, Gnanamanoharan E.

Computer Assisted Methods in Engineering and Science

|

2023

|

Vol. 30, no. 3

305--321

EN

Brain tumors are fatal for majority of the patients, the different nature of the tumorcells requires the use of combined medical measures, and categorizing such tumors isa difficult task for radiologists. The diagnostic structures based on PCs have been offeredas an aid in diagnosing a brain tumor using magnetic resonance imaging (MRI). Generalfunctions are retrieved from the lowest layers of the neural network, and these lowestlayers are responsible for capturing low-level features and patterns in the raw input data,which can be particularly unique to the raw image. To validate this, the EfficientNetB3pre-trained model is utilized to classify three types of brain tumors: glioma, meningioma,and pituitary tumor. Initially, the characteristics of several EfficientNet modules are takenfrom the pre-trained EfficientNetB3 version to locate the brain tumor. Three types of braintumor datasets are used to assess each approach. Compared to the existing deep learningmodels, the concatenated functions of EfficientNetB3 and genetic algorithms give betteraccuracy. Tensor flow 2 and Nesterov-accelerated adaptive moment estimation (Nadam)are also employed to improve the model training process by making it quicker and better.The proposed technique using CNN attains an accuracy of 99.56%, a sensitivity of 98.9%,a specificity of 98.6%, an F-score of 98.9%, a precision of 98.9%, and a recall of 99.54%.

8

Improved efficient capsule network for Kuzushiji-MNIST benchmark dataset classification

Bukowski Michał, Antoniuk Izabella, Kurek Jarosław

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2023

|

Vol. 71, nr 6

art. no. e147338

EN

In this paper, we present an improved efficient capsule network (CN) model for the classification of the Kuzushiji-MNIST and Kuzushiji-49 benchmark datasets. CNs are a promising approach in the field of deep learning, offering advantages such as robustness, better generalization, and a simpler network structure compared to traditional convolutional neural networks (CNNs). Proposed model, based on the Efficient CapsNet architecture, incorporates the self-attention routing mechanism, resulting in improved efficiency and reduced parameter count. The experiments conducted on the Kuzushiji-MNIST and Kuzushiji-49 datasets demonstrate that the model achieves competitive performance, ranking within the top ten solutions for both benchmarks. Despite using significantly fewer parameters compared to higher-rated competitors, presented model achieves comparable accuracy, with overall differences of only 0.91% and 1.97% for the Kuzushiji-MNIST and Kuzushiji- 49 datasets, respectively. Furthermore, the training time required to achieve these results is substantially reduced, enabling training on nonspecialized workstations. The proposed novelties of capsule architecture, including the integration of the self-attention mechanism and the efficient network structure, contribute to the improved efficiency and performance of presented model. These findings highlight the potential of CNs as a more efficient and effective approach for character classification tasks, with broader applications in various domains.

9

A deep learning method for hard-hat-wearing detection based on head center localization

Wójcik Bartosz, Żarski Mateusz, Książek Kamil, Miszczak Jarosław A., Skibniewski Mirosław J.

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2023

|

Vol. 71, nr 6

art. no. e147340

EN

In recent years, a lot of attention has been paid to deep learning methods in the context of vision-based construction site safety systems. However, there is still more to be done to establish the relationship between supervised construction workers and their essential personal protective equipment, like hard hats. A deep learning method combining object detection, head center localization, and simple rule-based reasoning is proposed in this article. In tests, this solution surpassed the previous methods based on the relative bounding box position of different instances and direct detection of hard hat wearers and non-wearers. Achieving MS COCO style overall AP of 67.5% compared to 66.4% and 66.3% achieved by the approaches mentioned above, with class-specific AP for hard hat non-wearers of 64.1% compared to 63.0% and 60.3%. The results show that using deep learning methods with a humanly interpretable rule-based algorithm is better suited for detecting hard hat non-wearers.

10

Random noise suppression of seismic data through multi-scale residual dense network

Gao Lei, Zhao KeCe, Min Fan, Bai Si-Chang

Acta Geophysica

|

2023

|

Vol. 71, no. 2

637--647

EN

Random noise suppression is an important technique to improve the efficiency and accuracy of seismic data processing. Physical denoising methods such as f − x deconvolution and K-SVD have been widely adopted by the industry, while popular learning-based methods such as neural networks have emerged as good alternatives. In this paper, we propose a multi-scale residual dense network (MSRDN) for random noise suppression of seismic raw data. First, the network consists of a shallow feature extraction module, multiple residual blocks and multiple up-sampling modules. They are used for feature extraction, noise learning and size restoration. Second, each residual block is composed of multiple dense blocks. They are designed to alleviate network degradation. Third, dense blocks are tightly connected by multi-scale convolutional layers. They can enhance the regularization effect of the network. The experimental results show that MSRDN is more accurate and stable than previous algorithms.

11

Interpretability versus explainability: classification for understanding deep learning systems and models

Namatēvs Ivars, Sudars Kaspars, Dobrājs Artis

Computer Assisted Methods in Engineering and Science

|

2022

|

Vol. 29, no. 4

297--356

EN

The techniques of explainability and interpretability are not alternatives for many realworld problems, as recent studies often suggest. Interpretable machine learning is nota subset of explainable artificial intelligence or vice versa. While the former aims to build glass-box predictive models, the latter seeks to understand a black box using an explanatory model, a surrogate model, an attribution approach, relevance importance, or other statistics. There is concern that definitions, approaches, and methods do not match, leading to the inconsistent classification of deep learning systems and models for interpretation and explanation. In this paper, we attempt to systematically evaluate and classify the various basic methods of interpretability and explainability used in the field of deep learning.One goal of this paper is to provide specific definitions for interpretability and explainability in Deep Learning. Another goal is to spell out the various research methods for interpretability and explainability through the lens of the literature to create a systematic classifier for interpretability and explainability in deep learning. We present a classifier that summarizes the basic techniques and methods of explainability and interpretability models. The evaluation of the classifier provides insights into the challenges of developinga complete and unified deep learning framework for interpretability and explainability concepts, approaches, and techniques.

12

The automatic focus segmentation of multi-focus image fusion

Hawari K., Ismail

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2022

|

Vol. 70, nr 1

art. no. e140352

EN

Multi-focus image fusion is a method of increasing the image quality and preventing image redundancy. It is utilized in many fields such as medical diagnostic, surveillance, and remote sensing. There are various algorithms available nowadays. However, a common problem is still there, i.e. the method is not sufficient to handle the ghost effect and unpredicted noises. Computational intelligence has developed quickly over recent decades, followed by the rapid development of multi-focus image fusion. The proposed method is multi-focus image fusion based on an automatic encoder-decoder algorithm. It uses deeplabV3+ architecture. During the training process, it uses a multi-focus dataset and ground truth. Then, the model of the network is constructed through the training process. This model was adopted in the testing process of sets to predict the focus map. The testing process is semantic focus processing. Lastly, the fusion process involves a focus map and multi-focus images to configure the fused image. The results show that the fused images do not contain any ghost effects or any unpredicted tiny objects. The assessment metric of the proposed method uses two aspects. The first is the accuracy of predicting a focus map, the second is an objective assessment of the fused image such as mutual information, SSIM, and PSNR indexes. They show a high score of precision and recall. In addition, the indexes of SSIM, PSNR, and mutual information are high. The proposed method also has more stable performance compared with other methods. Finally, the Resnet50 model algorithm in multi-focus image fusion can handle the ghost effect problem well.

13

Single target tracking algorithm for lightweight Siamese networks based on global attention

Wang Zhentao, He Xiaowei, Cheng Rao

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2022

|

Vol. 70, nr 3

art. no. e139961

EN

Object tracking based on Siamese networks has achieved great success in recent years, but increasingly advanced trackers are also becoming cumbersome, which will severely limit deployment on resource-constrained devices. To solve the above problems, we designed a network with the same or higher tracking performance as other lightweight models based on the SiamFC lightweight tracking model. At the same time, for the problems that the SiamFC tracking network is poor in processing similar semantic information, deformation, illumination change, and scale change, we propose a global attention module and different scale training and testing strategies to solve them. To verify the effectiveness of the proposed algorithm, this paper has done comparative experiments on the ILSVRC, OTB100, VOT2018 datasets. The experimental results show that the method proposed in this paper can significantly improve the performance of the benchmark algorithm.

14

MFFNet: A multi-frequency feature extraction and fusion network for visual processing

Deng Jinsheng, Zhang Zhichao, Yin Xiaoqing

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2022

|

Vol. 70, nr 3

art. no. e140466

EN

Convolutional neural networks have achieved tremendous success in the areas of image processing and computer vision. However, they experience problems with low-frequency information such as semantic and category content and background color, and high-frequency information such as edge and structure. We propose an efficient and accurate deep learning framework called the multi-frequency feature extraction and fusion network (MFFNet) to perform image processing tasks such as deblurring. MFFNet is aided by edge and attention modules to restore high-frequency information and overcomes the multiscale parameter problem and the low-efficiency issue of recurrent architectures. It handles information from multiple paths and extracts features such as edges, colors, positions, and differences. Then, edge detectors and attention modules are aggregated into units to refine and learn knowledge, and efficient multi-learning features are fused into a final perception result. Experimental results indicate that the proposed framework achieves state-of-the-art deblurring performance on benchmark datasets.

15

Hierarchical Bi-LSTM based emotion analysis of textual data

Mahto Dashrath, Yadav Subhash Chandra

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2022

|

Vol. 70, nr 3

art. no. e141001

EN

Nowadays, Twitter is one of the most popular microblogging sites that is generating a massive amount of textual data. Such textual data is intended to incorporate human feelings and opinions with related events like tweets, posts, and status updates. It then becomes difficult to identify and classify the emotions from the tweets due to their restricted word length and data diversity. In contrast, emotion analysis identifies and classifies different emotions based on the text data generated from social media platforms. The underlying work anticipates an efficient category and prediction technique for analyzing different emotions from textual data collected from Twitter. The proposed research work deliberates an enhanced deep neural network (EDNN) based hierarchical Bi-LSTM model for emotion analysis from textual data; that classifies the six emotions mainly sadness, love, joy, surprise, fear, and anger. Furthermore, the emotion analysis result obtained by the proposed hierarchical Bi-LSTM model is being compared and validated with the traditional hybrid CNN-LSTM approach regarding the accuracy, recall, precision, and F1-Score. It can be observed from the results that the proposed hierarchical Bi-LSTM achieves an average accuracy of 89% for emotion analysis, whereas the existing CNN-LSTM model achieved an overall accuracy of 75%. This result shows that the proposed hierarchical Bi-LSTM approach achieves desired performance compared to the CNN-LSTM model.

16

Traffic flow prediction in inland waterways of Assam region using uncertain spatiotemporal correlative features

Muthukumaran Venkatesan, Natarajan Rajesh, Kaladevi Amarakundhi Chandrasekaran, Magesh Gopu, Babu Swapna

Acta Geophysica

|

2022

|

Vol. 70, no 6

2979--2990

EN

Modern civilization has reported a significant rise in the volume of traffic on inland rivers all over the globe. Traffic flow prediction is essential for a good travel experience, but adequate computer processes for processing unpredictable spatiotemporal data (timestamp, weather, vessel_ID, water level, vessel_position, vessel_speed) in the inland water transportation industry are lacking. Moreover, such type of prediction relies primarily on past traffic patterns and perhaps other pertinent facts. Thus, we propose a deep learning-based computing process, namely Convolution Neural Network-Long Short-Term Memory Network (CNN-LSTM), a progressive predictor of employing uncertain spatiotemporal information to decrease navigation mishaps, traffic and flow prediction failures during transportation. Spatiotemporal correlation of current traffic flow may be processed using a simplified CNN-LSTM model. This hybridized prediction technique decreases update costs and meets the prediction needs with minimal computing overhead. A short case study on the waterways of the Indian state of Assam from Sandiya (27.835090 latitude, 95.658590 longitude) to Dhubri (26.022699 latitude, 89.978401 longitude) is undertaken to assess the model's performance. The evaluation of the suggested method includes a variety of trajectories of water transportation vehicles, including ferries, sailing boats, container ships, etc. The suggested approach outperforms conventional traffic flow predicting methods when it comes to short-term prediction with minimal predictive error (<2.75) and exhibited a major difference of more than 45% on the comparison of other methods.

17

The hydrological impact of tropical cyclones on soil moisture using a sensor based hybrid deep learning model

Srivastava Gautam, Kavitha V., Vimaladevi M., Parthiban N.

Acta Geophysica

|

2022

|

Vol. 70, no 6

2933--2951

EN

Tropical cyclones that originate from the Indian Ocean affect the Indian Sub-Continent. Heavy rainfall and flooding occur because of these cyclones. South Odisha was affected by Cyclonic Storm Daye in September 2018 and Cyclonic Storm Titli was occurred in August affecting Andhra Pradesh and Odisha as well. The Eastern portion of India was affected by the Cyclonic Storm Fani in April 2019. In May 2020, West Bengal was affected by the Amphan which is a Super Cyclonic Storm and in the same year Tamil Nadu was affected by the very severe Cyclonic Storm Nivar in November 2020. These are just a few of the notable cyclonic events in the Indian Sub-Continent. These cyclonic events cause a dramatic change in a very short time from dry soil to exceptional flooding. In this proposed work, we are attempting to create an observations-driven prediction model to quantify the soil moisture variations daily, predict county-based meteorology and evaluate the cause of cyclones and heavy rainfall in certain areas of India. In our work, we applied a deep learning-based methodology to predict soil moisture. For the prediction model, we fused Feed Forward Neural Networks with the Gated Recurrent Unit (GRU) model and present the prediction results. We have used climatic as well as environmental data published by the Indian Meteorological Department (IMD) Warning from 2011. The collected data is time-series data. Comparisons and the relationship that exists between soil moisture and meteorological data are made and analyzed. The soil moisture of the South Indian states Karnataka, Andhra Pradesh and Tamil Nadu are predicted from weather data using a hybrid deep learning model. The evaluations of the proposed work using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE) and R-squared (R2 ) against Non-hybrid Neural Network models such as Artificial Neural Networks (ANN), Convolutional Neural Networks, and Gated Recurrent Unit (GRU) models is analyzes where our model has given better results.

18

An automated prediction of remote sensing data of Queensland-Australia for food and wildfire susceptibility using BISSOA DBMLA scheme

Sankaran Krishnan Sakthidasan, Lim Se-Jung, Bhaskar Seelam Ch Vijaya

Acta Geophysica

|

2022

|

Vol. 70, no 6

3005--3021

EN

In past few decades, there was a tremendous enhancement in natural disaster and their effects on economy and population. An adverse events like foods, wildfires, cyclones, earthquakes, tsunamis, etc., are regarded as a natural disaster once it strikes the vulnerable population areas. An early tracking of susceptibility areas and immediate tracking of affected areas might help in facilitating rescue and early warnings to the public. To achieve autonomous natural disaster prediction, this paper makes use of current developments in remote sensing, which speed up the availability of aerial/satellite data and are reinforced by progress in the computing sector. The aerial/satellite imageries are employed for acquiring the data from areas of Queensland-Australia that are more prone to natural disaster in an eagle-eye perspective. Since there were several techniques employed so far for the automatic prediction of natural disaster susceptibilities, there were some limitations like reduced rate of accuracy and so on. So as to overcome these limitations, deep learning based automated process is employed for predicting the natural disaster areas and probability of event occurrences. The main intention of the work is to detect the natural disaster occurrence from the sensed data which aids in providing warning to public and to safeguard them by taking necessary actions. Initially, remote sensing data is pre-processed and the features are extracted using Adaptive linear Internal embedding algorithm-based feature extraction (ALIE-FE). The extracted features are selected using Recursive Wrapper-based feature subset selection. To estimate best fitness function and to enhance the prediction accuracy, the optimization process is carried using Bio-Inspired Squirrel Search Optimization algorithm (BI-SSOA). Finally, the classification is carried by means of Deep learning based Multi-layer Alex Net classifier (DBMLA) approach. The simulation is carried and the outcomes attained are estimated for predicting food susceptibility and wildfire susceptibility. The proposed BISSOA-DBMLA offers sensitivity of 98%, specificity of 99%, and TSS of 97%. The proposed system offers 98.99% classification accuracy. Accuracy, sensitivity, specificity, TSS, and area under the curve (AUC) are used to evaluate the efficacy of the suggested system in light of the achieved results from other approaches. To demonstrate the efficacy of the suggested mechanism, the achieved results are compared with those of current approaches.

19

2D inversion of magnetotelluric data using deep learning technology

Liao Xiaolong, Shi Zeyu, Zhang Zhihou, Yan Qixiang, Liu Pengfei

Acta Geophysica

|

2022

|

Vol. 70, no 3

1047--1060

EN

The inverse problem of magnetotelluric data is extremely difficult due to its nonlinear and ill-posed nature. The existing gradient-descent approaches for this task surface from the problems of falling into local minima and relying on reliable initial models, while statistical-based methods are computationally expensive. Inspired by the excellent nonlinear mapping ability of deep learning, in this study, we present a novel magnetotelluric inversion method based on fully convolutional networks. This approach directly builds an end-to-end mapping from apparent resistivity and phase data to resistivity anomaly model. The implementation of the proposed method contains two stages: training and testing. During the training stage, the weight sharing mechanism of fully convolutional network is considered, and only the single anomalous body model samples are used for training, which greatly shortens the modeling time and reduces the difficulty of network training. After that, the unknown combinatorial anomaly model can be reconstructed from the magnetotelluric data using the trained network. The proposed method is tested in both synthetic and field data. The results show that the deep learning-based inversion method proposed in this paper is computationally efficient and has high imaging accuracy.

20

Segmentation of bone structures with the use of deep learning techniques

Krawczyk Zuzanna, Starzyński Jacek

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2021

|

Vol. 69, nr 3

art. no. e136751

EN

The paper is focused on automatic segmentation task of bone structures out of CT data series of pelvic region. The authors trained and compared four different models of deep neural networks (FCN, PSPNet, U-net and Segnet) to perform the segmentation task of three following classes: background, patient outline and bones. The mean and class-wise Intersection over Union (IoU), Dice coefficient and pixel accuracy measures were evaluated for each network outcome. In the initial phase all of the networks were trained for 10 epochs. The most exact segmentation results were obtained with the use of U-net model, with mean IoU value equal to 93.2%. The results where further outperformed with the U-net model modification with ResNet50 model used as the encoder, trained by 30 epochs, which obtained following result: mIoU measure – 96.92%, “bone” class IoU – 92.87%, mDice coefficient – 98.41%, mDice coefficient for “bone” – 96.31%, mAccuracy – 99.85% and Accuracy for “bone” class – 99.92%.