Search

The Magnetic Observatory of Tatuoca (TTB) was installed by Observatório Nacional (ON) in 1957, near Belém city in the state of Pará, Brazilian Amazon. Its history goes back to 1933, when a Danish mission used this location to collect data, due to its privileged position near the terrestrial equator. Between 1957 and 2007, TTB produced 18,000 magnetograms on paper using photographic variometers, and other associated documents like absolute value forms and yearbooks. Data was obtained manually from these graphs with rulers and grids, taking 24 average readings per day, that is, one per hour. In 2017, the Federal University of Pará (UFPA in the Portuguese acronym) and ON collaborated to rescue this physical archive. In 2022 UFPA took a step forward and proposed not only digitizing the documents but also developing an intelligent agent capable of reading and extracting the information of the curves with a resolution better than an hour, being this the central goal of the project. If the project succeeds, it will rescue 50 years of data imprisoned in paper, increasing measurement sensitivity far beyond what these sources used to give. This will also open the possibility of applying the same AI to similar documents in other observatories or disciplines like seismography. This article recaps the project, and the complex challenges faced in articulating Archival Science principles with AI and Geoscience.

The purpose of this work is to explore the applicability and effectiveness of deep learning methods for the task------moves recognition in abstract of research paper. We firstly build a large corpus for moves recognition. Then we choose the traditional machine learning method SVM as a benchmark, and develop four moves recognition methods based on DNN, LSTM, Attention-BiLSTM and BERT. Finally, we design two groups of experiments with sample size 10,000 and 50,000 and then compare experimental results. The results show that most of the deep learning methods outperform the traditional machine learning method SVM especially in large-scale sample experiments, in which the BERT with a re-pre-trained model achieves the best results in both groups of experiments. Deep learning methods are proved applicable and effective for moves recognition in research paper abstracts.

To improve the accuracy of user implicit rating prediction, we combine the traditional latent factor model (LFM) and bidirectional gated recurrent unit neural network (BiGRU) model to propose a hybrid model that deeply mines the latent semantics in the unstructured content of the text and generates a more accurate rating matrix. First, we utilize the user’s historical behavior (favorites records) to build a user rating matrix and decompose the matrix to obtain the latent factor vectors of users and literature. We also apply the BERT model for word embedding of the research papers to obtain the sequence of word vectors. Then, we apply the BiGRU with the user attention mechanism to mine the research paper textual content and to generate the new literature latent feature vectors that are used to replace the original literature latent factor vectors decomposed from the rating matrix. Finally, a new rating matrix is generated to obtain users’ ratings of noninteractive research papers and to generate the recommendation list according to the user latent factor vector. We design experiments on the real datasets and verify that the research paper recommendation model is superior to traditional recommendation models in terms of precision, recall, F1-value, coverage, popularity and diversity.

Even current OCR technology has made significant progress with the help of deep learning, rapid identification algorithms for specific information in documents still need elaborate design. In this paper, a lightweight operating framework based on deep neural network is established to deal with the extraction of date and time information from paper worksheets used in transformer station. In the framework, YOLOv4-Tiny is employed to detect the locations of the dates and the time in document images which are mixed with printing and handwritten fonts. Since the handwritten dates in pictures are difficult to segment, a novel lightweight CNN named MobileDateNet is proposed that can recognize the date serially without segmentation. The experimental results show that the identification accuracy of the whole framework reaches 92.1% at the average running speed of 36.5 FPS on an NVIDIA 3090 GPU, ensuring the recognition accuracy and the operating efficiency.

The experiment of quality assurance of polymeric outdoor insulators service life is necessary to be checked the class of wettability by using the digital image processing tools. The image processing can improve its pictorial information for human interpretation, render it more suitable for autonomous machine perception. This paper provides a survey of the recent technology and theoretical concept to explain the development of the Hydrophobicity Class (HC) of porcelain insulator. The hydrophobic properties of the housing material were found degraded to a different extent between field-aged insulators due to differences in material structure and pollution conditions. This paper contributes to the latest development of reviews relating to image processing for the implementation of porcelain insulator hydrophobic systems testing., and their related studies. We categorized the computer vision mainstream into 2 group such as Machine Learning Methods Applied to Digital Image Processing and Deep Learning Methods Applied to Digital Image Processing. Then compare pros/cons in each type of technique. We also provide brief explanation on the up-to-date information about the techniques and their performance.

Lately, Deep Learning has indicated its capacity by adequately tackling complex learning issues which were not possible previously. Specifically, Convolutional Neural Networks are most generally utilized and have indicated their viability in image detection and recognition problems. Convolutional Neural Networks(CNN) can be used to perform different kinds of tasks. However, to recognize even a single image, billions of calculations are required. For example, ResNet 50 takes 8 billion calculations just to figure out what's in a single image. So we need to find out ways to be able to run that superfast and that's why we need hardware accelerators. In this paper the basic information about Convolutional Neural Networks is provided along with the key operations involved and the brief idea about Field Programmable Gate Arrays(FPGA) is given which enable them to be used for accelerating the inference process of Convolutional Neural Networks. Various techniques which were employed previously for accelerating the Convolutional Neural Networks are discussed. We are focussing on designing a energy efficient Field Programmable Gate Arrays accelerator for accelerating the inference process of Convolutional Neural Networks with reduced latency and increased throughput.

Due to the increasingly growing demand for user identification on cell phones, PCs, laptops, and so on, face anti-spoofing has risen to significance and is an active research area in academia and industry. The detection of the real face then recognize it present an important challenge regarding the techniques that can be used to spoof any recognition system like masks, printed photos. This paper we present an anti-spoofing face method to solve the real-world scenario that learns the target domain classifier based on samples used for training in a particular source domain. Specifically, with the conventional regression CNN, the Spatial/Channel-wise Attention Modules were introduced. Two modules, namely the Spatial-wise Attention Module and the Channel-wise Attention Module, were used at spatial and channel levels to improve local features and ignore the irrelevant features. Extensive experiments on current collections with benchmarks datasets verifies that the recommended solution will significantly benefit from the two modules and better generalization capability by providing significantly improved results in anti-spoofing.

Drowsiness can be dangerous when performing tasks that require constant attention, such as driving a vehicle. Sleepiness is correlated with a variety of physiological variables, such as eye closing, head movements, pulse rate, eye twitch rate, etc. Also, the yawn can be considered as an accurate indicator of drowsiness and fatigue. Yawning detection is very important for the safety purpose of drivers as it will let the driver know if he/she is getting drowsy. Driving at that moment may not be safe. Several automatic yawning detection techniques have been developed for driver's drowsiness monitoring system. Nevertheless, correctly detecting the yawning of the driver and predicting exhaustion in real-time situations is still a crucial challenge. In this paper, we will be reviewing various existing machine learning approaches for driver's yawning detection. In previous approaches, various classical machine learning algorithms such as viola-Jones, contour activation algorithm and SVM have been used for yawning detection, but these approaches failed to predict yawning in realtime situations. Using Deep learning techniques, we can make a real-time yawn detection system with high accuracy. We find that some precious Deep learning algorithms like CNN, RNN, LSTM, Bi-LSTM can detect the patterns with high accuracy. After the comparison of various algorithms and techniques, we find that with the help of Deep learning algorithms the yawning can be detected in real time with high accuracy.

Nowadays, there is a rapidly increasing number of conferences and journals in computer science that make a lot of challenges for researchers to find an appropriate venue to submit their scientific work. There is a need for a recommendation system that can support researchers for a better process of paper submission. In this paper, we present an efficient approach for building such a recommendation model by using embedding methods, Global Vector (GloVe) 1 created by Pennington et al. [1] and FastText 2 proposed by Facebook [2], Convolutional Neural Network (CNN) [3], and LSTM. The experimental results show that the combination of CNNs and FastText, CNN + FastText, can achieve the best performance in terms of the Top 1 Accuracy compared with other techniques, including the S2RSCS model, as presented in [4]. Moreover, the performance by using GloVe or FastText is much better, faster, and more stable than S2RSCS in most cases.

With the advances in information technologies, different research areas are emerging day by day and thousands of research papers are published in these fields. Papers are not presented to users grouped according to their topics. Therefore, it is getting harder and harder for people to access research papers in their field of interest. To facilitate this search process, a paper classification system is proposed in this study to categorize papers on similar topics. A different approach from the studies presented for classification so far has been proposed. This system is a complex system created by combining Word2vec, network modeling and community discovery. With the Word2vec method, which has attracted great attention recently, paper similarities have been found, a network based on similarity rates has been created and papers have been clustered with the community discovery algorithm. As a result of the application with the proposed system, a success of 89% has been achieved. As can be seen from the results, this approach presented for an important classification problem will provide great convenience to people. It will enable fast and efficient access to papers.

In the context of papers used in the graphic arts, including silver gelatin, inkjet, and wove papers, prior work has studied measures of texture similarity for purposes of classifying such papers. The majority of prior work has been based on classical image processing approaches such as Fourier, wavelet, and fractal analysis. In this work, recent advances in deep learning are used to develop a texture similarity approach for measuring paper texture similarity. Since the available datasets generally lack labels, the convolutional neural network is trained using triplet loss to minimize the feature distance of tiles from the same image while simultaneously maximizing the feature distance of tiles drawn from different images. The approach is tested on three paper texture image databases considered in prior works and the results suggest the proposed approach achieves state-of-the-art performance.

Which are the segmentation algorithms proposed during 2018-2019 in CVPR that have CNN architecture?’ Answering this question involves identifying and analyzing the deep learning architecture diagrams from several research papers. Retrieving such information poses significant challenge as most of the existing academic search engines are primarily based on only the text content. In this paper, we introduce Diag2Graph, an end-to-end framework for parsing deep learning diagram-figures, that enables powerful search and retrieval of architectural details in research papers. Our proposed approach automatically localizes figures from research papers, classifies them, and analyses the content of the diagram-figures. The key steps in analyzing the Figure content is the extraction of the different components data and finding their structural relation. Finally, the extracted components and their relations are represented in the form of a deep knowledge graph. A thorough evaluation on a real-word annotated dataset has been done to demonstrate the efficacy of our approach.

In recent years, the volume of research papers has become enormous. Therefore, it is difficult for researchers to select their required papers. To lighten this problem, a method of describing abstracts called "Structured Abstract" is used in the fields of physiology and medicine. This paper proposes a method that extracts sentences matching with each heading of Structured Abstract. Specifically, SciBERT and BioBERT, which are specialized in the science and technology fields, are used in the method. The effectiveness of the proposed method is evaluated by comparing automatically generated abstracts and human-written abstracts.

Prior work on texture analysis of historic, photographic papers has focused primarily on measures of texture similarity. However, automated grouping or clustering of photographic paper textures in a way that is meaningful to art conservators remains an open problem. In this work a deep learning approach to automated classification is presented, for clusters derived from a human sorting experiment conducted by 19 art conservators and paper experts and subsequently extended through crowd-sourcing. The proposed approach uses a deep convolutional neural network, and results are presented on the performance in automatically classifying images when compared to human experts.

In autonomous driving, 3D object detection is es-sential as it provides basic knowledge about the environment. However, as deep learning based 3D detection methods are usually computation intensive, it is challenging to support realtime 3D object detection on edge-computing devices in selfdriving cars with limited computation and memory resources. To facilitate this, we propose a compiler-aware pruning search framework, to achieve real-time inference of 3D object detection on the resource-limited mobile devices. Specifically, a generator is applied to sample better pruning proposals in the search space based on current proposals with their performance, and an evaluator is adopted to evaluate the sampled pruning proposal performance. To accelerate the search, the evaluator employs Bayesian optimization with an ensemble of neural predictors. We demonstrate in experiments that for the first time, the pruning search framework can achieve real-time 3D object detection on mobile (Samsung Galaxy S20 phone) with state-of-the-art detection performance.

We investigate end-to-end optimized optical transmission systems based on feedforward or bidirectional recurrent neural networks (BRNN) and deep learning. In particular, we report the first experimental demonstration of a BRNN auto-encoder, highlighting the performance improvement achieved with recurrent processing for communication over dispersive nonlinear channels.

Steganography is the art of hiding a secret message inside a publicly visible carrier message. Ideally, it is done without modifying the carrier, and with minimal loss of information in the secret message. Recently, various deep learning based approaches to steganography have been applied to different message types. We propose a deep learning based technique to hide a source RGB image message inside finite length speech segments without perceptual loss. To achieve this, we train three neural networks; an encoding network to hide the message in the carrier, a decoding network to reconstruct the message from the carrier and an additional image enhancer network to further improve the reconstructed message. We also discuss future improvements to the algorithm proposed.

In recent times, Generative Adversarial Networks (GANs) have created a lot of buzz in the research community. GANs are formulated on the zero-sum game theory, where two neural nets compete against each other. The resultant deep model is capable of generating data similar to any data distribution provided. It utilizes the adversarial learning approach and is far more capable in learning features than the traditional machine learning models. This review focusses on the origin and evolution of GANs. Firstly, the traditional GAN is explored in terms of its structure and loss functions. Then come the common challenges of training GANs. Thirdly, the review dives into numerous GAN variants and explains their improvements. The review then lists the wide variety of applications and ends with the conclusion.

In the graph classification problem, given is a family of graphs and a group of different categories, and we aim to classify all the graphs (of the family) into the given categories. Earlier approaches, such as graph kernels and graph embedding techniques have focused on extracting certain features by processing the entire graph. However, real world graphs are complex and noisy and these traditional approaches are computationally intensive. With the introduction of the deep learning framework, there have been numerous attempts to create more efficient classification approaches. We modify a kernel graph convolutional neural network approach, that extracts subgraphs (patches) from the graph using various community detection algorithms. These patches are provided as input to a graph kernel and max pooling is applied. We use different community detection algorithms and a shortest path graph kernel and compare their efficiency and performance. In this paper we compare three methods: a graph kernel, an embedding technique and one that uses convolutional neural networks by using eight real world datasets, ranging from biological to social networks.

Electrical vehicular (EV) energy management is a promising trend. Forecasting vehicular trajectories and delay is crucial for EV energy management. The presented work is devoted to the study and the application of deep learning techniques on specific road trajectories. First, exhaustive deep learning algorithms are considered. Second, road traces are converted to time series. Then, delays and road trajectories are analyzed. In fact, we consider two Recurrent Neural Networks (RNN): LSTM (Long Short Term Memory) and GRU (Gated Recurrent Units). Neural Networks are adapted and trained on 60 days of real urban traffic of Rome in Italy. We calculate the Loss function for both machine learning techniques which is defined by mean square error (MSE) and Root mean square error (RMSE). Experimental results demonstrate that both LSTM and GRU are adequate for the context of EV in terms of route trajectory and delay prediction.

In this paper, we investigate the fast downlink beamforming for full-dimension multiple-input multiple-output systems under correlated Rician channels. Under the assumption that the base station (BS) has only statistical channel state information (CSI), we decouple each user’s beamforming vector and derive their optimal beamforming vector through the maximization of the average signal-to-leakage-plus-noise ratio (SLNR) lower bound. Then, to reduce the computation time, a model-driven deep learning (DL)-based beamforming algorithm is proposed, as well as a data-driven algoriothm for comparison. In the model-driven DL-based beamforming algorithm, the process of obtaining the beamforming vector is separated into two parallel neural networks which are constructed and trained independently. The proposed algorithms can achieve similar ergodic rate as the optimal beamforming algorithm with much less computation time, and the model-driven algorithm requires less computing resource than the data-driven algorithm.

Breast cancer diagnosis requires a pathologist to analyze the histology slides under various magnifications. An automated diagnosis method to aid pathologists that is magnification independent will significantly save time, reduce cost and mitigate subjectivity and errors in current histopathological diagnosis procedures. This paper presents a deep learning network, called MVPNet and a customized data augmentation technique, called NuView, for magnification independent diagnosis. MVPNet is tailored to tackle the most common issues (diversity, relatively small size of datasets and manifestation of diagnostic biomarkers at various magnification levels) with breast cancer histology data to perform the classification. The network simultaneously analyzes local and global features of a given tissue image. It does so by viewing the tissue at varying levels of relative nuclei sizes. MVPNet has significantly less parameters than standard transfer learning deep models with comparable performance and it combines and processes local and global features simulatenously for effective diagnosis. Additionally, NuView extracts tumor nuclei location and points the attention of MVPNet to the informative region specifically. The method gives an average magnification independent classification accuracy of 92.2% as compared to 83% reported in literature on the BreaKHis database.

In today's open source era, developers look forsimilar software applications in source code repositories for anumber of reasons, including, exploring alternative implementations, reusing source code, or looking for a better application. However, while there are a great many studies for finding similarapplications written in the same programming language, there isa marked lack of studies for finding similar software applicationswritten in different languages. In this paper, we fill the gapby proposing a novel modelCroLSimwhich is able to detectsimilar software applications across different programming lan-guages. In our approach, we use the API documentation tofind relationships among the API calls used by the differentprogramming languages. We adopt a deep learning based word-vector learning method to identify semantic relationships amongthe API documentation which we then use to detect cross-language similar software applications. For evaluating CroLSim, we formed a repository consisting of 8,956 Java, 7,658 C#, and 10,232 Python applications collected from GitHub. Weobserved thatCroLSimcan successfully detect similar softwareapplications across different programming languages with a meanaverage precision rate of 0.65, an average confidence rate of3.6 (out of 5) with 75% high rated successful queries, whichoutperforms all related existing approaches with a significantperformance improvement.

In the new era of genetic epidemiology, there have been growing interest in studying genetic variants and their associations to complex diseases. Advances in modern computational approaches have led to the search for useful interacting genetic variants that are associated to the manifestation of a disease. However, these conventional strategies face number of challenges in predicting interesting interactions when data acquisition and dimensionality increases. Deep learning promises empirical success in number of applications including bioinformatics to drive insights of biological complexities. A deep neural network was previously proposed to identify true causative two-locus SNP interactions. The method was evaluated on various simulated and real datasets. In this study, the performance of the previously proposed deep learning method is maximized by improving network learning and avoiding overfitting. The method is further extended for performing sensitivity analysis. The performance of the method is evaluated on chronical dialysis patient's data for identifying higher-order interactions. It was observed that the highly ranked two-locus and three-locus SNP interactions in mitochondrial D-loop has the highest risk for the manifestation of disease.

Electrophoresis (EP) test separates protein components based on their density. Patterns exhibited by this test mostly show very close approximation, making it difficult to examine test results within a short amount of time as it has many variations of patterns and requires a significant amount of knowledge to discern them accurately. To help clinical examiners save time and produce consistent results, a new deeplearning model optimized for EP graphic images was developed. Extending recent work on capsule network, which is a stateof- the-art deep learning model, this study was carried out to develop a best-performing model in classifying abnormal and normal electrophoresis patterns. Instead of extracting features from the image, we used the whole slide image as an input to the classifier. This study used 39,484 electrophoresis 2D graph images and utilized capsule network as the foundation of the deep learning architecture to learn the images without data augmentation. The formulated models were compared for a multitude of performance metrics including accuracy, sensitivity, and specificity. Overall, the study results show that our proposed architecture EP-CapsNet, which combines capsule network with Google’s inception module, is the best performing model, outperforming the baseline and alternative models in almost all comparisons.

In this Plenary paper the most important scientific challenge of knowledge learning is reviewed thought two different approaches: Deep Learning (DL) and Wise Learning (WL).

When studying metabolic disease, it is essential to investigate the disease's effect on multiple tissues and identify any communication, or cross-talk, between organs, tissues, and cells. In bone marrow cancer, adipose tissue triggers inflammation and growth of malignant plasma cells within the bone marrow and results in localized bone loss. Synchrotron radiation microtomography imaging enables 3D quantitative analysis of bone and adipose tissues and provides high resolution to observe local changes in tissue microstructure. However, optimal imaging techniques differ for hard bone tissues (absorption imaging) and soft adipose tissues (phase-contrast imaging). Here we introduce a new technique that leverages image reconstruction and deep learning in combination with the high-resolution imaging capabilities of synchrotron radiation microtomography to gain insight into the marrow microenvironment of human bone samples. This approach allowed for successful tissue segmentation and analysis of human core samples. Using high-resolution images such as these could allow for a better understanding of early bone-related changes that may predict disease progression or bone fractures.