Search

Global supply chains are kept viable through the information shared through billions of electronic documents, many of which extensively use tables to display critical information. Making effective supply chain decisions requires the extraction of data from these tables which is hindered by the variations in layouts and styles of tables. In this paper, we propose Table Det: a deep learning based methodology to solve table detection and table image classification in data sheet images in a single inference as the first stage of the table text extraction pipeline. TableDet utilizes Cascade R-CNN with Complete IOU (CIOU) loss and a deformable convolution backbone as its underlying architecture to capture the variations in scales and orientations of tables. It also detects text and figures to enhance its table detection performance. We demonstrate the effectiveness of training TableDet with a dual-step transfer learning process and fine-tuning it with Table Aware Cutout (TAC) augmentation strategy. We achieved the highest F1 score for table detection against state-of-the-art solutions on ICDAR 2013 (complete set), ICDAR 2017 (test set) and ICDAR 2019 (test set) with 100%, 99.3% and 95.1% respectively. For the table image classification task we attained 100% recall and above 85% precision on three test sets. This classification capability ensures that all images with tables would be promoted to the next step in the table text extraction pipeline, with a small number of images without tables making it through.

There is an increasing demand for automated document processing techniques as the volume of electronic component documents increase. This is most prevalent in the supply chain optimization sector where vast amount of documents need to be processed and is time consuming and prone to error. Detection of tables and table structures serves as a crucial step to automate document processing. While table detection is a well investigated problem, tabular structure detection is more complex, and requires further improvements. To address this, this study proposes a deep learning model that focuses on high precision tabular cell structure detection. The proposed model creates a benchmark for the ICDAR2013 dataset cell structure with comparison to the previous state of the art table detection models as well as proposing alternative models. Our methodology approaches improving table structure detection through the detection of cells instead of row and columns for better generalization capabilities for heterogeneous table structures. Our proposed model advances prior models by improving major parts of the detection pipeline, mainly the two-stage detector, backbone, backbone architecture, and non-maximum-suppression (NMS). TabCellNet consists of Hybrid Task Cascade (HTC) with Combinational Backbone Network (CBNet), dual ResNeXt101 and Soft-NMS to achieve a precision of 89.2% and recall of 98.7% on the hand annotated ICDAR2013 cell structure dataset.

Because it contains high spectral information, hyperspectral imagery has been used in many areas. However, hyperspectral imagery has low spatial resolution because of imaging hardware limitation. Recently, many methods have been available for improving spatial resolution of hyperspectral images. Pan-sharpening and dictionary learning-based sparse representation methods are well-known methods for improving spatial resolution. In this study, a quantitative analysis of super-resolution methods for hyperspectral imagery is performed for identifying the best method in terms of reconstruction quality and processing time. K-SVD, ODL and Bayesian methods are employed for dictionary learning-based sparse representations. On the other hand, IHS and PCA-based methods are employed for pan-sharpening methods. The experimental results show that the ODL method outperforms others in terms of reconstruction quality measured by RMSE values and processing times.

The nondedicated sensing capabilities of smart mobile devices contribute to Internet of Things (IoT) ecosystems with integral building blocks called mobile crowdsensing (MCS) systems. The distributed and nontrusted nature of MCS systems leads to various threats for devices and MCS platforms as well as for end users. Out of the many threats, fake tasks may lead to drained resources at the participating devices and clogged resources at the MCS platforms. Furthermore, when limited data are available, it becomes a further challenge to identify maliciously submitted fake tasks. In this article, we introduce possible solutions that leverage ensemble learning against fake tasks submitted to MCS platforms. More specifically, boosting-based solutions, namely adaptive boosting for binary classification (AdaBoost), gentle adaptive boosting (GentleBoost), and random under-sampling boosting (RUSBoost), form the basis for learning the legitimacy of tasks submitted to MCS platforms. Over a six-day observation window, one day was used for training while the remaining five days were used for testing to evaluate the performance under limited data in terms of training the machine learning (ML) models. Through extensive simulations, we have shown that GentleBoostbased ensemble learning can achieve promising performance in detecting fake/illegitimate tasks submitted to an MCS platform.

Mobile crowdsensing (MCS) is a distributed sensing concept that enables ubiquitous sensing services via various builtin sensors in smart devices. However, MCS systems are vulnerable because of being non-dedicated. Especially, submission of fake tasks with the aim of clogging participants device resources as well as MCS servers is a crucial threat to MCS platforms. In this paper, we propose an ensemble learning-based solution for MCS platforms to mitigate illegitimate tasks. Furthermore, we also integrate k-means-based classification with the proposed method to extract region-specific features as input to the machine learningbased fake task detection. Through simulations, we compare the ensemble method to a previously proposed Deep Belief Network (DBN)-based fake task detection, which is also shown to improve performance in terms of accuracy, F1 score, recall, precision and geometric mean score (G-mean) with the integration of regionawareness. Our validation results show that the ensemble machine learning-based detection can eliminate majority of the fake tasks, with up to 0.995 precision, 0.997 recall, 0.996 F1, 0.993 accuracy and 0.982 G-Mean. Furthermore, the proposed solution introduces savings up to 12.18% battery of mobile devices while reducing the impacted recruits to 0.25% and protecting up to 10.59% participants against malicious sensing tasks.

Mobile Crowdsensing (MCS) builds on the Sensing as a Service model, and is considered to be an integral component of the Internet of Things systems. Since MCS does not build on a thoroughly assessed and established trust mechanism between all parties various threats including data poisoning, fake sensing tasks and clogging task attacks remain challenges. Fake task submissions are the least investigated although they have the potential to drain significant amount of resources (e.g. battery, sensors, processing, storage) and clog the MCS servers. This paper proposes a knowledge-based technique alongside sequential feature selection methodology to detect fake sensing tasks submitted to the MCS servers so that the tasks do not get assigned to the participants but filtered at the MCS servers. The proposed methodology is compared to fake task detection under Knowledge Based Deep Neural Network which is also enhanced by feature selection, and the simulation results show that the proposed methodology, by utilizing Deep Prior Knowledge Input with Self-Organizing Feature Map can outperform the deep neural network-based detection by 9.7% in terms of accuracy.

Connected vehicle networks and future autonomous driving systems call for characterization of risky behavior to improve safety models and autonomous driving features. While risky behavior patterns entail potential safety issues on road networks, the advent of vehicular sensing and vehicular networks cannot guarantee accurate characterization of driving/movement behavior of vehicles. One of the roadblocks against robust event characterization systems in connected vehicles is the scarcity of anomalous data to enable training of event classification models. With this in mind, the contribution of this paper is two-fold: 1) a reliable methodology to generate representative data under the scarcity of diverse anomalous sensory data, 2) classification of mobility/driving events of vehicles with high accuracy. To this end, as a baseline method, an optimized deep recurrent neural network-based encoding model is introduced to extract the precise feature representation of the anomalous data. In addition to this, an event characterization pipeline is introduced that uses the representation provided by the encoder to generate reliable data, then train and classify future events. To improve the classification performance of the baseline method, a long short-term memory (LSTM)-based auto-encoder network is proposed. Through experimental results, it is shown that the LSTM-based auto-encoder network can achieve over 0.93 accuracy, which outperforms the baseline recurrent neural network model by 12%.

Documents are constantly being processed within supply chains in various industries throughout the globe. Within those documents, often times the most important content is stored in tabular format. Therefore an automated technique for supply chain document processing is highly desired. Deep learning approaches show promise to deliver an end-to-end extraction model. However, it has been shown that tabular detection accuracy is not always correlated to tabular localization accuracy. Portions of the desired tabular information can easily be cropped out due to a lack of localization accuracy. In this paper, we propose a two stage convolutional neural network-based deep learning framework to improve tabular localization accuracy. We use pre-trained backbone network ResNet-50 and then apply transfer learning to fit our application. One of our main contributions is the introduction of the KL loss function into Faster-RCNN. Once the bounding box variances are acquired from the KL loss function, we introduce a voting procedure with soft-non-maximum suppression (Soft-NMS) to improve localization performance. The proposed framework is trained and evaluated on public and private datasets that span from scientific documents to various electronic components. Our test results show that the precision of tabular detection can be improved by 1.2% while achieving the same recall as other state-of-the-art models on the public ICDAR2013 dataset. Furthermore, a large improvement in precision has been achieved at extremely high intersection over union (IoU) thresholds (i.e. 95%). Thus, 10.9% higher precision is achieved at 95% IoU for ICDAR2013. For another public dataset, namely ICDAR2017, 8.4% higher precision is achieved at 95% IoU .

Mobile crowdsensing (MCS) is a ubiquitous sensing paradigm that emerged in the form of”sensed data as a service” model in the Internet of Things Era. Distributed nature of MCS results in vulnerabilities at the MCS platforms as well as participating devices that provide sensory data services. Submission of fake tasks with the aim of clogging sensing server resources and draining participating device batteries is a crucial threat that has not been investigated well. In this paper, we provide a detailed analysis by modeling a deep belief network (DBN) when the available sensory data is scarce for analysis. With oversampling to cope with the class imbalance challenge, a Principal Component Analysis (PCA) module is implemented prior to the DBN and weights of various features of sensing tasks are analyzed under varying inputs. The experimental results show that the presented DBN-driven fake task mitigation detection of fake sensing tasks can ensure up to 0.92 accuracy, 0.943 precision and up to 0.928 F1 score outperforming prior work on MCS data with deep learning networks.

The advent of artificial intelligence methodologies pave the way towards smarter healthcare by exploiting new concepts such as deep learning. This chapter presents an overview of deep learning techniques that are applied to smart healthcare. Deep learning techniques are frequently applied to smart health to enable AI-based recent technological development to healthcare. Furthermore, the chapter also introduces challenges and opportunities in deep learning particularly in the healthcare domain.

Citizen-centric methods leverage connectivity and Artificial Intelligence (AI) methodologies to improve smart city services. Among these services, intelligent re-design of first responder services call for novel solutions that save city resources and result in faster response time. To this end, we propose to utilize sensory data in connected vehicles to capture contextual details that can be obtained through various Convolutional Neural Network (CNN) architectures to determine which set of first responders should be called in the case of an accident. We use real images from a rich dataset of accidents involving different types of vehicles to train and test the CNNs. Through experimental results, we show that when ResNet-34 network is augmented with one fit cycle, image augmentation and hidden layer unfreezing methods can result in 88.9% accuracy in the prediction of the required first responder(s) in case of an accident solely based on the captured images.

Mobile crowdsensing (MCS) is a ubiquitous sensing paradigm where built-in sensors of smart mobile devices are empowered to acquire sensory data in lieu of dedicating large scale sensing infrastructures. One of the most crucial problems in mobile crowdsensing is the injection of fake sensing tasks to clog the energy, computing, storage and sensing resources of participating devices. In this paper, we present solutions that leverage deep networks to analyze the tasks submitted to MCS platforms. To this end, we model off-the-shelf deep learning models, namely Deep Autoencoder (Deep-AE), Restricted Boltzmann Machine (RBM) and Deep Belief Network (DBN) in order to detect and filter out illegitimate tasks submitted to MCS campaigns. For the same purpose, we also utilize a Deep Multi-layer Perceptron (Deep-MLP) network instead of the well known Multi-layer Perceptron. Through numerical results on MCS data, we show that Deep-MLP outperforms its counterparts with 0.963 precision and 0.964 recall in the detection of fake sensing tasks.

Mobile crowdsensing (MCS) is a cloud-inspired and non-dedicated sensing paradigm to enable ubiquitous sensing via built-in sensors of personalized devices. Due to disparate participants and sensing tasks, MCS is vulnerable to threats initiated by malicious participants, which can either be a participant providing sensory data or an end user injecting a fake task aiming at resource (e.g. battery, sensor, etc.) clogging at the participating devices. This paper builds on machine learning-based detection of illegitimate tasks, and investigates the impact of machine learning-based prevention of battery-oriented illegitimate task injection in MCS campaigns. To this end, we introduce two different attack strategies, and test the impact of ML-based detection and elimination of fake tasks on task completion rate, as well as the overall battery drain of participating devices. Simulation results confirm that up to 14% battery power can be saved at the expense of a slight decrease in the completion rate of legitimate tasks.

Location-based clogging attacks in a Mobile Crowdsensing (MCS) system occur following upon the submission of fake tasks, and aim to consume the batteries and hardware resources of smart mobile devices such as sensors, memory and processors. Intelligent modeling of fake task submissions is required to enable the development of effective defense mechanisms against location-based clogging attacks with fake task submissions. An intelligent strategy for fake task submission would aim to maximize the impact on the participants of an MCS system. With this in mind, this paper introduces new algorithms exploiting the Self-Organizing Feature Map (SOFM) to identify attack locations where fake sensing tasks submitted to an MCS platform are centered around. The proposed SOFM-based model addresses issues in the previously proposed SOFM-based attack models by proposing two ways of refinement. When compared to the former models, which also use SOFM architectures, simulation results show that up to 139.9% of impact improvement can be modeled under the reconfigurable SOFM architectures.

The global outbreak of the Coronavirus Disease 2019 (COVID-19) pandemic has uncovered the fragility of healthcare and public health preparedness and planning against epidemics/pandemics. In addition to the medical practice for treatment and immunization, it is vital to have a thorough understanding of community spread phenomena as related research reports 17.9&ndash, 30.8% confirmed cases to remain asymptomatic. Therefore, an effective assessment strategy is vital to maximize tested population in a short amount of time. This article proposes an Artificial Intelligence (AI)-driven mobilization strategy for mobile assessment agents for epidemics/pandemics. To this end, a self-organizing feature map (SOFM) is trained by using data acquired from past mobile crowdsensing (MCS) campaigns to model mobility patterns of individuals in multiple districts of a city so to maximize the assessed population with minimum agents in the shortest possible time. Through simulation results for a real street map on a mobile crowdsensing simulator and considering the worst case analysis, it is shown that on the 15th day following the first confirmed case in the city under the risk of community spread, AI-enabled mobilization of assessment centers can reduce the unassessed population size down to one fourth of the unassessed population under the case when assessment agents are randomly deployed over the entire city.

Extracting data from tabular structures contained within product datasheets is crucial in many contexts, particularly in the management and optimization of supply chains that serve various industries. In order to minimize human intervention, table detection and table structure detection form the essential functionality. However, a self-contained holistic solution to extract the tables as well as their columns and rows in not readily available. To address this challenge, This study presents a new formal procedure that consists of the following sequence: table detection, structure segmentation and holistic tabular structure detection on documents. The proposed table detection model outperforms the state-of-the-art solutions by achieving a recall value of 1.0 and a precision of more than 0.99 on public competition datasets. Furthermore, this work introduces a judging mechanism and an agreement-based post-processing procedure to incorporate hand-crafted rules into the deep learning models. Though the individual components achieve a new state-of-the-art F1-Score, when integrated the best achieved F-measure for the holistic system is 0.89.

Artificial neural networks have been recognized as an important technique in microwave modeling and optimization. This paper gives an overview and recent developments on the knowledge based neural modeling techniques in microwave modeling and design. The knowledge based artificial neural networks are constructed by incorporating the existing knowledge such as empirical formulas, equivalent circuit models and semi-analytical equations in neural network structures. The existing knowledge reduces the complexity of neural network model. This combination requires less training data and has better extrapolation performance than classical neural networks. The advantages of using knowledge based neural network modeling are demonstrated with two microwave modeling applications such as characteristic impedance modeling of thin-film microstrip line and parametric modeling of the differential via holes.

SUMMARY This paper presents two new methods, space mapping (SM) with prior knowledge input (PKI-D) with difference and compound space mapping-based neuromodeling. Both methods combine two powerful techniques, space mapping-based neuromodeling and PKI-D with difference. The knowledge-based modeling methods in the RF/microwave literature merge the prior knowledge about the device to be modeled with neural network structures while a knowledge-based method, SP, focuses on reducing the computational burden. The main advantage of the proposed methods over these already existing knowledge-based methods are their better extrapolation capability and reduced number of training set data. The simulation results obtained reveal that both methods decrease the cost of training and improve the extrapolation capability and output performance of the SP-based neuromodeling. Copyright # 2007 John Wiley & Sons, Ltd.

International Conference Information Communication Automation Technologies (ICAT) -- OCT 29-31, 2015 -- Sarajevo, BOSNIA & HERCEG WOS: 000380438700014 The spatial resolutions of hyperspectral images are generally lower due to imaging hardware limitations. Super-resolution algorithms can be applied to obtain higher resolutions. Many algorithms exist to achieve super-resolution hyperspectral images from low resolution images acquired in different wavelengths. One of the popular algorithms is sparse representation-based algorithms that employ dictionary learning methods. In this study, a comparative framework is developed to investigate which dictionary learning algorithm leads to better super-resolution images. In order to achieve that, K-SVD and ODL dictionary learning algorithms are employed for comparison. A sparse representation-based algorithm G-SOMP+ is used for hyperspectral super-resolution reconstruction. The experimental results show that ODL algorithm outperforms K-SVD in terms of both reconstruction quality and processing times. Univ Sarajevo, Fac Elect Engn Sarajevo, IEEE, IEEE CSS, IEEE Comp Soc, IEEE SMC

Artificial Neural Network (ANN) is an important technique for modeling and optimization in engineering design. It is very suitable in modeling as it needs only the information based on relationship between the input and the output related to the problem. For further improvement in modeling, a priori knowledge about the problem such as an empirical formula, an equivalent circuit model, and a semi-analytical equation is directly embedded in ANN structure through a knowledge based modeling strategy. Three-step modeling strategy that exploits knowledge based techniques is developed to improve some properties of conventional ANN modeling such as accuracy and data requirement. All these improvements ensure better accuracy with less time consumption compared to conventional ANN modeling. The necessary knowledge in this strategy is generated in the first step through conventional ANN. Then this knowledge is embedded in the new ANN model for the second step. Final model is constructed by incorporating the existing knowledge obtained by the second step. Therefore each model generates better accuracy than previous model. Conventional ANN, prior knowledge input, and prior knowledge input with difference techniques are used to improve accuracy, time consumption, and data requirement of the modeling in three-step modeling strategy. The efficiency of three-step modeling strategy is demonstrated on the nonlinear function modeling and the high dimensional shape reconstruction problem.

Summary Artificial neural network (ANN) provides an efficient modeling technique based on input–output data obtained from an engineering problem. Highly nonlinear and complex relationships can be formed by ANN because of its nonlinear nature and representing knowledge at interconnection weights. If an ANN model is not sufficient in respect to the accuracy and time consumption, the knowledge based on design experience can be embedded into the modeling process. This knowledge reduces the complexity of the nonlinear input–output relationships; therefore, the knowledge embedded ANN can be formed easily compared with the conventional ANN model. Three-step modeling strategy generates the initial knowledge via the conventional ANN modeling and consists of three sequential and also discrete training processes exploiting the knowledge-based methods such as prior knowledge input and prior knowledge input with difference. The latter step improves the accuracy of the former step, and three-step modeling provides more accuracy than conventional ANN modeling. The efficiency of three-step modeling strategy is demonstrated on two data sets, which are obtained by the reconfigurable microstrip patch antenna design problem. The different number of neurons and ANN structures is handled for the comparison as well. In addition, ANN modeling is formed by MATLAB m-file through neural network toolbox to reveal the efficiency of knowledge-based three-step modeling strategy. Copyright © 2016 John Wiley & Sons, Ltd.

Artificial neural networks have been used as an important technique in modeling and optimization for engineering design. In this work, 3-step modeling strategy based on knowledge based techniques is proposed to develop new efficient modeling instead of conventional artificial neural network (ANN) modeling. The knowledge based artificial neural networks are constructed by incorporating the existing knowledge such as empirical formulas, equivalent circuit models and semi-analytical equations in neural network structures. In this new technique, required knowledge is created in the first step and used in the second step as a coarse model. Therefore each model shows better performance than former. In this strategy, conventional ANN, prior knowledge input and prior knowledge input with difference techniques are utilized not only to improve modeling accuracy but also to reduce time consumption during modeling. The advantages of using 3-step modeling are demonstrated on Branin function modeling application.

Artificial Neural Networks (ANN) have emerged as a powerful technique for modeling. Since the embedding knowledge in ANN models is possible by the Knowledge Based ANN (KBANN) methods, more accurate results than classical ANN approach can be obtained with KBANN. Source Difference (SD), Prior Knowledge Input (PKI) and Prior Knowledge Input with Difference (PKI-D) are several methods to be mentioned which combines existing knowledge with ANN methods. The existing knowledge is obtained either by mathematical formulations, ANN modeling or measured data. The Prior Knowledge Input with Difference, which is the latest method amongst KBANN approaches is discussed in this work. We compared the response efficiency and time consumption performances of PKI-D and classical ANN methods to obtain model for Inverse Scattering Problem.

Artificial neural networks have been used as an important technique in microwave modelling and optimisation. This paper gives an overview and recent developments on the knowledge-based neural modelling techniques in microwave modelling and design. The knowledge-based artificial neural networks are constructed by incorporating the existing knowledge such as empirical formulas, equivalent circuit models and semi-analytical equations in neural network structures. When one of the knowledge-based methods can not provide sufficient accuracy, two of them can be used in the same modelling process. This combination of methods is named hybrid technique. Using knowledge-based techniques requires less training data and has better extrapolation performance than classical neural networks. The advantages of using knowledge-based neural network modelling are demonstrated with microwave device modelling applications.

24th Signal Processing and Communication Application Conference (SIU) -- MAY 16-19, 2016 -- Zonguldak, TURKEY WOS: 000391250900166 Due to hardware limitations, hyperspectral imagery has low spatial resolution. It can be obtained super-resolution hyperspectral imagery by means of sparse representation-based methods that are designed for improving spatial resolution. In this paper, the effect of sparse representation-based dictionary learning algorithms including K-SVD, ODL and Bayes on obtaining superresolution images with low error and high quality has been investigated. The method with best results has been identified. IEEE, Bulent Ecevit Univ, Dept Elect & Elect Engn, Bulent Ecevit Univ, Dept Biomed Engn, Bulent Ecevit Univ, Dept Comp Engn