Search

Technology is shaping our lives in a multitude of ways. This is fuelled by a technology infrastructure, both legacy and state of the art, composed of a heterogeneous group of hardware, software, services, and organisations. Such infrastructure faces a diverse range of challenges to its operations that include security, privacy, resilience, and quality of services. Among these, cybersecurity and privacy are taking the centre-stage, especially since the General Data Protection Regulation (GDPR) came into effect. Traditional security and privacy techniques are overstretched and adversarial actors have evolved to design exploitation techniques that circumvent protection. With the ever-increasing complexity of technology infrastructure, security and privacy-preservation specialists have started to look for adaptable and flexible protection methods that can evolve (potentially autonomously) as the adversarial actor changes its techniques. For this, Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) were put forward as saviours. In this paper, we look at the promises of AI, ML, and DL stated in academic and industrial literature and evaluate how realistic they are. We also put forward potential challenges a DL based security and privacy protection system has to overcome. Finally, we conclude the paper with a discussion on what steps the DL and the security and privacy-preservation community have to take to ensure that DL is not just going to be hype, but an opportunity to build a secure, reliable, and trusted technology infrastructure on which we can rely on for so much in our lives.

The rapid growth of scientific literature is presenting several challenges for the search and discovery of research artifacts. Datasets are the backbone of scientific experiments. It is crucial to locate the datasets used or generated by previous research as building suitable datasets is costly in terms of time, money, and human labor. Hence automated mechanisms to aid the search and discovery of datasets from scientific publications can aid reproducibility and reusability of these valuable scientific artifacts. Here in this work, utilizing the next sentence prediction capability of language models, we show that a BERT-based entity recognition model with POS aware embedding can be effectively used to address this problem. Our investigation shows that identifying sentences containing dataset mentions in the first place proves critical to the task. Our method outperforms earlier ones and achieves an F1 score of 56.2 in extracting dataset mentions from research papers on a popular corpus of social science publications. We make our codes available at https://github.com/sandeep82945/data_discovery.

This paper presents the development of a test paper checker that will recognize a handwritten text using Intelligent Character Recognition (ICR) for Alphanumeric Characters. An examination can be conducted in two ways—digital and manual—and each way has a different approach in checking. In this study, the main objective is to recognize alphanumeric handwritten characters accurately using intelligent character recognition. OpenCV is used in the Python programming language and Support Vector Machine as a tool in machine learning for ICR. Answer sheet was designed with 120 items for MCQ and problem-solving questions. Item analysis and printing of results are included in the device. Experiments were conducted by giving an actual examination from the 131 participants in Technological University of the Philippines for testing the accuracy of the device. The results obtained from comparing manual and machine checking had an accuracy of 93.0769%. Thus, the proposed method is applicable for the development of handwritten character recognition.

E-commerce customers and general Web users should not believe that the products suggested by Recommender systems or results displayed by Web search engines are either complete or relevant to their search aspirations. The economic priority of Web related businesses requires a higher rank on Web snippets or product suggestions in order to receive additional customers; furthermore, Web search engines and recommender systems revenue is obtained from advertisements and pay-per-click. This survey paper presents a review of Web Search Engines, Ranking Algorithms, Citation Analysis and Recommender Systems. In addition, Neural networks and Deep Learning are also analyzed including their use in learning relevance and ranking. Finally, this survey paper also introduces the Random Neural Network with its practical applications.

Thousands of academic papers are submitted at top venues each year. Manual audits are time-consuming and laborious. And the result may be influenced by human factors. This paper investigates a modularized and attention-based recurrent convolutional network model to represent academic paper and predict aspect scores. This model treats input text as module-document hierarchies, uses attention pooling CNN and LSTM to represent text, and outputs prediction with a linear layer. Empirical results on PeerRead data show that this model give the best performance among the baseline models.

The citation similarity measurement task is defined as determining how similar the meanings of two citations are. This task play an significant role in Natural Language Processing applications, especially in academic plagiarism detection. Yet, computing citation similarity is not a trivial task, due to the incomplete and ambiguous information presented in academic papers, which makes necessity to leverage extra knowledge to understand it, as well as most similarity measures based on the syntactic features, and other based on the semantic part still has many drawbacks. In this paper, we propose a corpus-based approach using deep learning word embeddings to compute more effective citation similarity. Our study explores the previous works on text similarity, namely, string-based, knowledge-based and corpus-based. Then we define our new basis and experiment on a large dataset of scientific papers. The final results demonstrate that deep learning based approach can enhance the effectiveness of citation similarity.

In recent years, a lot of vehicle attacks have been reported and demonstrated by researchers and whitehat hackers indicating vehicle cyber security as an important issue particularly for self-driving cars. The reason behind this extended attack vector is the multiple external interfaces of vehicles and minimal internal security protection. Hence, it is totally possible for adversaries to take full control of connected cars. In this paper, we propose an efficient Vehicular Intrusion Detection System (IDS), named as VIDS, which consists of a lightweight domain-based detection model for ECU devices and a comprehensive crossdomain-based detection model for a gateway or domain controller. The former makes use of specification periodic features of Controller Area Network (CAN) frames, while the latter exploits stream bit value features with deep learning techniques. With the use of real vehicular normal datasets and synthesized abnormal datasets for experimenting, the experimental results indicate that the proposed VIDS can achieve better detection rate over existing IDS systems. In addition, VIDS is compatible with vehicle internal CAN network.

COVID-19 is an infectious disease, which has adversely affected public health and the economy across the world. On account of the highly infectious nature of the disease, rapid automated diagnosis of COVID-19 is urgently needed. A few recent findings suggest that chest X-rays and CT scans can be used by machine learning for the diagnosis of COVID-19. Herein, we employed semi-supervised learning (SSL) approaches to detect COVID-19 cases accurately by analyzing digital chest X-rays and CT scans. On a relatively small COVID-19 radiography dataset, which contains only 219 COVID-19 positive images, 1341 normal and 1345 viral pneumonia images, our algorithm, COVIDCon, which takes advantage of data augmentation, consistency regularization, and multicontrastive learning, attains 97.07% average class prediction accuracy, with 1000 labeled images, which is 7.65% better than the next best SSL method, virtual adversarial training. COVIDCon performs even better on a larger COVID-19 CT Scan dataset that contains 82,767 images. It achieved an excellent accuracy of 99.13%, at 20,000 labels, which is 6.45% better than the next best pseudo-labeling approach. COVIDCon outperforms other state-of-the-art algorithms at every label that we have investigated. These results demonstrate COVIDCon as the benchmark SSL algorithm for potential diagnosis of COVID-19 from chest X-rays and CT-Scans. Furthermore, COVIDCon performs exceptionally well in identifying COVID-19 positive cases from a completely unseen repository with a confirmed COVID-19 case history. COVIDCon, may provide a fast, accurate, and reliable method for screening COVID-19 patients.

In recent years, AI based on deep learning has achieved tremendous success in specialized tasks such as speech recognition, machine translation, and the detection of tumors in medical images. Despite these successes there are also some clear signs of the limitations of the current state-of-the-art in AI. For example, biases in AI-enabled face recognition and predictive policing have shown that prejudice in AI systems is a real problem that must be solved. In this position paper, we argue that current AI needs to be developed along four dimensions to become more generally applicable and trustworthy: environment, purpose, collaboration, and governance. Hybrid AI offers the potential for advancements along these four dimensions by combining two different paradigms in AI: knowledge-based reasoning and optimization, and data-driven machine learning

Image registration is one of the most challenging problems in medical image analysis. In the recent years, deep learning based approaches became quite popular, providing fast and performing registration strategies. In this short paper, we summarise our work presented on Learn2Reg challenge 2020. The main contributions of our work rely on (i) a symmetric formulation, predicting the transformations from source to target and from target to source simultaneously, enforcing the trained representations to be similar and (ii) integration of variety of publicly available datasets used both for pretraining and for augmenting segmentation labels. Our method reports a mean dice of 0.64 for task 3 and 0.85 for task 4 on the test sets, taking third place on the challenge. Our code and models are publicly available at https://github.com/TheoEst/abdominal_registration and https://github.com/TheoEst/hippocampus_registration.

Visual odometry can be used to estimate the pose of a robot from current and recent video frames. A problem with these methods is that they drift over time due to the accumulation of estimation errors at each time-step. In this short paper we propose and briefly demonstrate the potential benefit of using prior 2D, top-down map information combined with multiple hypothesis particle filtering to correct visual odometry estimates. The results demonstrate a substantial improvement in robustness and accuracy over the sole use of visual odometry.

Nowadays, Machine Learning techniques are more and more pervasive in several application fields. In order to perform an evaluation as reliable as possible, it is necessary to consider the reproducibility of these models both at training and inference time. With the introduction of Deep Learning (DL), the assessment of reproducibility became a critical issue due to heuristic considerations made at training time that, although improving the optimization performances of such complex models, can result in non-deterministic outcomes and, therefore, not reproducible models. The aim of this paper is to quantitatively highlight the reproducibility problem of DL approaches, proposing to overcome it by using statistical considerations. We show that, even if the models generated by using several times the same data show differences in the inference phase, the obtained results are not statistically different. In particular, this short paper analyzes, as a case study, our ICPR2018 DL based approach for the breast segmentation in DCE-MRI, demonstrating the reproducibility of the reported results.

Transfer learning using deep pre-trained convolutional neural networks is increasingly used to solve a large number of problems in the medical field. In spite of being trained using images with entirely different domain, these networks are flexible to adapt to solve a problem in a different domain too. Transfer learning involves fine-tuning a pre-trained network with optimal values of hyperparameters such as learning rate, batch size, and number of training epochs. The process of training the network identifies the relevant features for solving a specific problem. Adapting the pre-trained network to solve a different problem requires fine-tuning until relevant features are obtained. This is facilitated through the use of large number of filters present in the convolutional layers of pre-trained network. A very few features out of these features are useful for solving the problem in a different domain, while others are irrelevant, use of which may only reduce the efficacy of the network. However, by minimizing the number of filters required to solve the problem, the efficiency of the training the network can be improved. In this study, we consider identification of relevant filters using the pre-trained networks namely AlexNet and VGG-16 net to detect cervical cancer from cervix images. This paper presents a novel hybrid transfer learning technique, in which a CNN is built and trained from scratch, with initial weights of only those filters which were identified as relevant using AlexNet and VGG-16 net. This study used 2198 cervix images with 1090 belonging to negative class and 1108 to positive class. Our experiment using hybrid transfer learning achieved an accuracy of 91.46%.

Machine learning has several potential uses in medical imaging for semantic labeling of images to improve radiologist workflow and to triage studies for review. The purpose of this study was to (1) develop deep convolutional neural networks (DCNNs) for automated classification of 2D mammography views, determination of breast laterality, and assessment and of breast tissue density; and (2) compare the performance of DCNNs on these tasks of varying complexity to each other. We obtained 3034 2D-mammographic images from the Digital Database for Screening Mammography, annotated with mammographic view, image laterality, and breast tissue density. These images were used to train a DCNN to classify images for these three tasks. The DCNN trained to classify mammographic view achieved receiver-operating-characteristic (ROC) area under the curve (AUC) of 1. The DCNN trained to classify breast image laterality initially misclassified right and left breasts (AUC 0.75); however, after discontinuing horizontal flips during data augmentation, AUC improved to 0.93 (p

The diagnosis of breast cancer in early stage is essential for successful treatment. Detection can be performed in several ways, the most common being through mammograms. The projections acquired by this type of examination are directly affected by the composition of the breast, which density can be similar to the suspicious masses, being a challenge the identification of malignant lesions. In this article, we propose a computer-aided detection (CAD) system to aid in the diagnosis of masses in digitized mammograms using a model based in the U-Net, allowing specialists to monitor the lesion over time. Unlike most of the studies, we propose the use of an entire base of digitized mammograms using normal, benign, and malignant cases. Our research is divided into four stages: (1) pre-processing, with the removal of irrelevant information, enhancement of the contrast of 7989 images of the Digital Database for Screening Mammography (DDSM), and obtaining regions of interest. (2) Data augmentation, with horizontal mirroring, zooming, and resizing of images; (3) training, with tests of six-based U-Net models, with different characteristics; (4) testing, evaluating four metrics, accuracy, sensitivity, specificity, and Dice Index. The tested models obtained different results regarding the assessed parameters. The best model achieved a sensitivity of 92.32%, specificity of 80.47%, accuracy of 85.95% Dice Index of 79.39%, and AUC of 86.40%. Even using a full base without case selection bias, the results obtained demonstrate that the use of a complete database can provide knowledge to the CAD expert.

The reconstruction quality of dental computed tomography (DCT) is vulnerable to metal implants because the presence of dense metallic objects causes beam hardening and streak artifacts in the reconstructed images. These metal artifacts degrade the images and decrease the clinical usefulness of DCT. Although interpolation-based metal artifact reduction (MAR) methods have been introduced, they may not be efficient in DCT because teeth as well as metallic objects have high X-ray attenuation. In this study, we investigated an effective MAR method based on a fully convolutional network (FCN) in both sinogram and image domains. The method consisted of three main steps: (1) segmentation of the metal trace, (2) FCN-based restoration in the sinogram domain, and (3) FCN-based restoration in image domain followed by metal insertion. We performed a computational simulation and an experiment to investigate the image quality and evaluated the effectiveness of the proposed method. The results of the proposed method were compared with those obtained by the normalized MAR method and the deep learning–based MAR algorithm in the sinogram domain with respect to the root-mean-square error and the structural similarity. Our results indicate that the proposed MAR method significantly reduced the presence of metal artifacts in DCT images and demonstrated better image performance than those of the other algorithms in reducing the streak artifacts without introducing any contrast anomaly.

Deep learning has demonstrated great success in various computer vision tasks. However, tracking the lumbar spine by digitalized video fluoroscopic imaging (DVFI), which can quantitatively analyze the motion mode of the spine to diagnose lumbar instability, has not yet been well developed due to the lack of steady and robust tracking method. The aim of this work is to automatically track lumbar vertebras with rotated bounding boxes in DVFI sequences. Instead of distinguishing vertebras using annotated lumbar images or sequences, we train a full-convolutional siamese neural network offline to learn generic image features with transfer learning. The siamese network is trained to learn a similarity function that compares the labeled target from the initial frame with the candidate patches from the current frame. The similarity function returns a high score if the two images depict the same object. Once learned, the similarity function is used to track a previously unseen object without any adapting online. Our tracker is performed by evaluating the candidate rotated patches sampled around the previous target's position and presents rotated bounding boxes to locate the lumbar spine from L1 to L4. Results indicate that the proposed tracking method can track the lumbar vertebra steadily and robustly. The study demonstrates that the lumbar tracker based on siamese convolutional network can be trained successfully without annotated lumbar sequences.

Many companies are studying autonomous vehicles. One trend in the development of control algorithms for autonomous vehicles is the use of deep-learning approaches. The general idea is to simulate a human driver’s decision-making and behavior in various scenarios without necessarily knowing why the decision is made. In this position paper, we first argue that traditional safety analysis methods need to be extended to verify deep-learning-based autonomous vehicles. Then, we propose borrowing ideas from the process of issuing driving licenses to human drivers to verify autonomous vehicles. Verification of autonomous vehicles could focus on sufficient training as well as mental and physical health checks. Based on this position, we list several challenges that need to be addressed.

This position paper discusses the definition and implementation of Web-Physical (WP) user profiles, which allow the creation of personalized recommendations and innovative behavioral predictions in particular scenarios, i.e., fairs. The nature of a WP profile builds upon two different worlds: the Web (social networks and web applications) and the Physical one, each one of them being explored through (big) data collection platforms. These two platforms collect radically different information: on the one hand, information of appreciation towards a particular product or service (web domain) together with other metadata; on the other, the leases (x, y) of users in the exhibition space (physical domain). In this scenario, our research idea consists in identifying how the information in the two domains can be merged in a whole entity under a theoretical point of view: this will unleash tangible repercussions in terms of personalized recommendations and effective behavioral predictions, where with personalized recommendation we mean a suggestion to a user in physical terms (eg a pavilion to visit) and / or in web terms (eg a site to visit) and with behavioral prediction a prediction of where a user can go in the future, even in a multimedia perspective (physical + web).

Nowadays the number of vehicles on road is increasing incrementally in emerging countries, and there is a need for proactive road health inspection methods that allow non-laborious and frequent road inspection with reasonable cost. Because of the nature of cheap and non-intrusive, vision-based road analysis has become a very popular topic in vision-based civil engineering researches around the world. In this paper, we discuss recent vision-based road distress detection methods, including our previous work and our presented work, for road distress visual inspection. This paper is divided into two parts. In the first part, we present non-deep learning based methods that aim to solve this problem of automatic road distress inspection. In the second part, we analyze recent deep learning based methods for road distress visual inspection and other related inspection of structure health monitoring. Finally, we propose our deep learning model called Modified U-Net whose goal is to solve the very difficult task of vision-based road’s crack detection. Along this paper, we also discuss about human factors in developing vision-based solutions for automatic road distress detection.

Deep neural networks (DNNs) are becoming widespread, and can often outperform manually-created systems. However, these networks are typically opaque to humans, and may demonstrate undesirable behavior in corner cases that were not encountered previously. In order to mitigate this risk, one approach calls for augmenting DNNs with hand-crafted override rules. These override rules serve to prevent the DNN from making certain decisions, when certain criteria are met. Here, we build on this approach and propose to bring together DNNs and the well-studied scenario-based modeling paradigm, by encoding override rules as simple and intuitive scenarios. We demonstrate that the scenario-based paradigm can render override rules more comprehensible to humans, while keeping them sufficiently powerful and expressive to increase the overall safety of the model. We propose a method for applying scenario-based modeling to this new setting, and apply it to multiple DNN models. (This paper substantially extends the paper titled “Guarded Deep Learning using Scenario-Based Modeling”, published in Modelsward 2020 [47]. Most notably, it includes an additional case study, extends the approach to recurrent neural networks, and discusses various aspects of the proposed paradigm more thoroughly).

Users’ opinions can be greatly beneficial in developing and providing products and services and improving marketing techniques for customer recommendation and retention. For this reason, sentiment analysis algorithms that automatically extract sentiment information from customers’ reviews are receiving growing attention from the computer science community. Aspect-based sentiment analysis (ABSA) allows for a more detailed understanding of customer opinions because it enables extracting sentiment polarities along with the sentiment target from sentences. ABSA consists of two steps: Aspect Extraction (AE) that allows recognizing the target sentiment; Aspect Sentiment Classification (ASC) that enables to classify the sentiment polarity. Currently, most diffused sentiment analysis algorithms are based on deep learning. Such algorithms require large labeled datasets that are extremely expensive and time consuming to build. In this paper, we present two approaches based on transfer learning and weak supervision, respectively. Both have the goal of reducing the manual effort needed to build annotated datasets for the ASC problem. In the paper, we describe the two approaches and experimentally compare them.

Most public domain images are exposed to copyright infringement due to different Rights Management Information (RMI) expression system and/or out-of-date RMI. To solve this problem, this paper proposes a reliable method of integrating the RMI representation system and updating the RMI with the up-to-date information based on the most reliable data among information collected from each site through a comparative search technique for public domain images based on deep learning. Through the RMI updating method proposed in this paper, public domain image sites will be able to minimize copyright infringement, since they can provide the most up-to-date RMI and may transform different RMI representation systems into a unified system.

The aim of the experiments described in this paper is to evaluate the use of keyboard and mouse dynamics as an appropriate non-obtrusive sensory input for an system that is sensitive to the affective state of its user. Our motivation for starting this research line has been the lack of tools and methodologies for taking into account this affective state in learning environments. In an ideal situation, when instructors have to choose from a collection of programming assignments, they should consider the studentś affective state and skills in order to select a learning resource with the appropriate level of difficulty. However, neither the data or the ability to process it are present in current learning management systems. This work tries to address this problem, by focused on the capture and pre-processing of data that is going to be fed to several machine learning techniques with the objective of classifying the affective states of users with different levels of expertise when learning a new programming language. We capture student data from a web-based platform, a learning management system where students interact with programming exercises. We introduce in this paper a series of pre-processing techniques that are able to convert data from keyboard and mouse dynamics captured from students as they were coding basic Python programming assignments into feature vector, that have been later used for the classification into five affective states: boredom, frustration, distraction, relaxation and engagement. The following classification algorithms have been evaluated: k-nearest neighbors,feed forward neural networks, naive Bayes classifier, J-48 tree induction algorithm, deep learning, random forest, gradient boosted trees and naive Bayes Kernel). The best accuracy was around 78% and was achieved by the tree induction algorithms. Results show that data gathered from ready-available, non-obtrusive sensors can be used successfully as another input to hybrid classification models in order to predict an individualś affective states, and these in turn can be used to make students more engaged and thus learn more efficiently.

Purpose Automatic identification of interventional devices in X-ray (XR) fluoroscopy offers the potential of improved navigation during transcatheter endovascular procedures. This paper presents a prototype implementation of fully automatic 3D reconstruction of a cryo-balloon catheter during pulmonary vein isolation (PVI) procedures by deep learning approaches. Methods We employ convolutional neural networks (CNN) to automatically identify the cryo-balloon XR marker and catheter shaft in 2D fluoroscopy during PVI. Training data are generated exploiting established semiautomatic techniques, including template-matching and analytical graph building. A first network of U-net architecture uses a single grayscale XR image as input and yields the mask of the XR marker. A second network of the similar architecture is trained using the mask of the XR marker as additional input to the grayscale XR image for the segmentation of the cryo-balloon catheter shaft mask. The structures automatically identified in two 2D images with different angulations are then used to reconstruct the cryo-balloon in 3D. Results Automatic identification of the XR marker was successful in 78% of test cases and in 100% for the catheter shaft. Training of the model for prediction of the XR marker mask was successful with 3426 training samples. Incorporation of the XR marker mask as additional input for the model predicting the catheter shaft allowed to achieve good training result with only 805 training samples. The average prediction time per frame was 14.47 ms for the XR marker and 78.22 ms for the catheter shaft. Localization accuracy for the XR marker yielded on average 1.52 pixels or 0.56 mm. Conclusions In this paper, we report a novel method for automatic detection and 3D reconstruction of the cryo-balloon catheter shaft and marker from 2D fluoroscopic images. Initial evaluation yields promising results thus indicating the high potential of CNNs as alternatives to the current state-of-the-art solutions.

The detection of roads from satellite images is among the most important topics for planning and development for cities, which replaces manual methods, but they turn out to be a complex task due to the complexity of the objects. This paper discusses issues related to the detection and segmentation of roads in very high-resolution aerial images. In order to resolve these issues, we propose in this paper the use, deployment, and validation of deep learning strategies, in particular, the U-net architecture based on deep convolutional neural networks for extracting roads from remote sensing images. Data augmentation techniques and preprocessing were applied to improve accuracy. We use in our processing for road segmentation the Massachusetts Road Dataset, which is publicly available. The results obtained showed excellent performance in terms of recall, precision, accuracy, and F1 score, and they are very close to the ground truth; it outperforms all other models presented, with a high accuracy of 97.7%.

Image Colorization is the problem of defining colors for grayscale images. Deep neural networks proved a great success in different fields recently. Therefore, it is used to solve the image colorization problem; moreover, it proved to be a very good choice. In literature, few review papers addressed the colorization problem. Most of them classified the research papers according to one or two criteria as input image type, a number of colored output images, colorization methodology, techniques or networks used in colorization, and network paths. This review classifies the papers according to these criteria intagrally and with a relatively large number of papers. Besides, the review displays the commonly used datasets and measures of comparison. It is found that deep learning is a widely used solution methodology to the problem. Unifying the comparison measures and data sets might help show the advances of the new models.

Diabetic retinopathy (DR) is a diabetes complication that damages the retina. This type of medical condition affects up to 80% of patients with diabetes for 10 or more years. The expertise and equipment required are often lacking in areas where diabetic retinopathy detection is most needed. Most of the work in the field of diabetic retinopathy has been based on disease detection or manual extraction of features. Thus, this research aims at automatic diagnosis of the disease in its different stages using deep learning neural network approach. This paper presents the design and implementation of Graphic Processing Unit (hereby GPU) accelerated deep convolutional neural networks to automatically diagnose and thereby classify high-resolution retinal images into five stages of the disease based on its severity. The accuracy of the single model convolutional neural networks presented in this paper is 71.65% from VGG-16.

Recently, Machine learning models gaining more popularity and can be considered as a significant investment and valuable assets for companies. These models can be used in a wide variety of tasks and in different fields of medicine, automation and robotics in factories, FinTech for banking and financial sectors as well as in modern e-commerce and marketing platforms. If these machine learning models are not well protected by intellectual property (IP), competitor can copy and steal these models and they can tweak these models to avoid detection with very little time and efforts and claiming these models as their own products. There are special considerations should be highlighted to properly draft and publish successful machine learning model patent. This paper aims to highlight the best practices of drafting patents of newly innovative machine learning models. Special focuses on the possible directions of these innovations where machine learning models can be considered as the modern solutions that need to be protected as valuable investments. This paper focused on the increased recent importance of artificial intelligence-related inventions and attempting to clarify doubts about AI-patents. The definition and understanding of functional claiming is critical step toward successful AI-patents. These efforts help researchers and developers to overcome the challenges of ambiguity of AI-patents as these patents totally different in aspects than traditional hardware circuitry patents for example. Therefore, with these guidance of the best practices the building blocks functions can be easily identified for innovative methods that can AI patentable.

An increasing volume of malicious spam calls has ushered an urgent need for a system that can automatically detect them. A robust spam call detection system must identify and block spam calls with high accuracy. Unlike current techniques that use crowd sourcing and are prone to phone number spoofing, this paper proposes a novel deep learning-based approach to classify phone calls, as spam or ham, based on the generated call transcript. This paper proposes a means of classification using embedded features with a 2-Path Hybrid Model (2PHM), comprising of a Long Short Term Memory Network (LSTM) and a Sentence Similarity Network. The 2PHM uses the transcript generated from the call by converting it into feature vectors, embedded in a high dimensional latent space, using the Universal Sentence Encoder embedding technique. Both paths of the hybrid model generate results that are weighed in an 85:15 ratio to finally generate a classification. The proposed system has been tested against Machine Learning models like Naive Bayes, K-Nearest Neighbors, Random Forests and Support Vector Machines and performs better than them, with an overall classification accuracy of 93% and a latency of 3 s.

Brand identity plays a vital role in business success. With the strong development of the market economy and scientific and technologiccal revolution, there are many new brands introduced to the market. So, how can customers identify that brand belongs to what industry or that logo is true or false? Therefore, the enterprises should have good strategy to enhance their competitiveness, especially in the recognizing brand. Logo of the enterprise can be used as suitable objects in computer vision applications for recognizing brands and providing associated services such as logo-based commercial research, and brand trend analysis. In this paper, we will present an overview of the brand and the importance of brand recognition. Then, the paper discusses the brand determination by using different approaches, thereby showing the pros and cons of these methods. Finally, we propose strategies to improve the competitiveness for enterprises in brand recognition based on machine learning model. The results show that our method increased the performance in brand recognition with large input size; conduce to help businesses maintain; and expand and improve trust for customers. It can also contribute to prevent unfair competition, and enhance the enterprise’s position in the domestic and international market.

Due to the complexity of the domestic road environment, the real-time detection and anti-collision warning of various categories of objects on the road are the key points to be solved. The safety assistant driving system based on machine vision is a good solution to avoid danger. Through the use of deep learning algorithm, it can detect the driving, pedestrians, non-motor vehicles and other real-time and anti-collision warning in the road environment. At the same time, it can detect whether drivers have bad driving behaviors such as fatigue driving, driving distraction and timely warning, so as to ensure the safety of the driving environment. This paper mainly studies the auxiliary driving safety early warning system based on the Internet of vehicles. According to the deep learning algorithm, this paper studies the target detection algorithm and collision early warning, analyzes and compares the principles of the related algorithms in machine vision ranging, and completes the camera calibration and the camera parameter acquisition required by visual distance measurement, through stereo matching and depth map mapping, target recognition and range detection are completed. In order to study the accuracy of the relative distance detection estimation algorithm of the camera, this paper selects 6 sample pictures in the special data training center of the road, and uses the distance detection and estimation algorithm to carry out the distance detection experiment, and calculates the distance of the detected object. The experimental results show that the accuracy of distance detection decreases with the increase of distance, the relationship between them is negative.

As a rapid, accurate, convenient, and non-destructive agricultural product analysis technology, NIR analysis technology has been widely used in agricultural product quality testing and raw material source identification. It is generally considered to be a non-destructive that is expected to further replace traditional agricultural chemical analysis detection technology. At present, the identification of citrus soil origin identification and crop quality monitoring technology based on near-infrared spectroscopy analysis is still relatively complicated, time-consuming, labor-consuming, and not accurate enough, and its completeness, systemicity and functionality are still not practical. The rapid identification of the origin of citrus and the effective quality supervision and inspection technology system play a very important role in promoting the healthy development of the citrus industry in our country. Based on this, this paper proposes a NIR-origin spectrum identification model based on a deep limit machine learning machine to further improve the accuracy of NIR-origin spectrum identification. First, we have a general understanding of the NIR origin spectrum identification to be studied in this article. To pave the way for the experimental part, and then propose an algorithm based on the depth limit machine (DELM) according to related technical issues, and verify it experimentally. The experimental results show that the DELM classification accuracy proposed in this paper reaches 0.932. The accuracy is better than the other two algorithms, but the PCA+machine Learning method is similar to the DELM effect, but the calculation is more complicated, requiring multiple steps such as dimensionality reduction+classifier construction, and DELM directly inputs the spectral information through depth the network and ELM are effectively integrated to construct an end-to-end feature extraction+deep classification model, which is more direct and efficient. In addition, compared with ELM, DELM can learn high-level and more abstract spectral feature expression more effectively, so it has a higher recognition rate.

With the continuous progress of science and technology, people get information on the Internet more and more quickly, the way is more and more convenient. People are more willing to use mobile devices, computers and other electronic products than to obtain information from paper books, newspapers and other channels. While using these electronic products, there will be many behaviors of storing or spreading bad information. Traditional OCR technology can be widely used in the detection and recognition of ordinary single background level text detection and recognition, but for the natural scene pictures with a certain Angle of the text detection and recognition effect is not good. With the development of deep learning technology, text detection and text recognition are mostly completed by deep learning architecture. In this paper, the popular EAST algorithm and CRNN algorithm are combined to detect and identify the text information in the natural scene.

Ship type identification in a maritime context can be critical to the authorities to control the activities being carried out. Although Automatic Identification Systems (AIS) has been mandatory for certain vessels if a vessel does not have them voluntarily or not, it can lead to a whole set of problems, so the use of tracking alternatives such as radar is fully complementary. However, radars provide positions, but not what they are detecting. Having systems capable of adding categorical information to radar detections of vessels makes it possible to increase control of the activities being carried out, improve safety in maritime traffic, and optimize on-site inspection resources on the part of the authorities. This paper addresses the binary classification problem (fishing ships versus all other vessels) using unbalanced data from real vessel trajectories. It is performed from a Deep Learning (DL) approach comparing two of the main trends, Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). In this paper it is proposed the weighted Cross-Entropy (WCE) methodology and compared with classical data balancing strategies. Both networks show high performance when applying WCE compared to the classical machine learning approaches and classical balancing techniques. This work is shown to be a novel approach to the international problem of identifying fishing ships without contexts.

Detection of COVID-19 by using patient’s blood platelets using an artificial neural network to perform large-scale automated testing. The deep learning model uses the CBC test results to detect the virus with an accuracy of 89.26%, which achieve higher accuracy than the widely used RT-PCR test. The model has achieved accuracy with only a thousand rows, which assures that the model can achieve better results with more data. The objective of the paper is to provide an automated and accurate method to detect COVID-19 using machine learning to apply it in large-scale testing, to help in containing the virus and prevent the spread of the virus. The scope of the paper is on the detection methodology using the blood platelets values extracted from the complete blood cycle (CBC) testing and discussing the future possibilities of improving the automated prevention method. The paper uses a dataset provided by Hospital Israelita Albert Einstein in Brazil, which has over 50 features of different tests as Influenza A, Influenza B, Uria, etc.…, many of these features had missing values or wasn’t relevant to the existence of the virus in the body;therefore, the feature selection led to filtering the 17 blood features. The K-Nearest Neighbour (KNN) Imputation imputed the missing values in the 17 filtered values. The analysis & visualization of the blood features concluded that the complete blood cycle test values highly correlated with each other, and there is a change occurs due to the presence of the virus is clear, especially to Eosinophils and Leukocytes, at which they represent the white blood cells forming the immune system that reacts against any virus or bacteria attracting the human body. The paper has applied five different machine learning algorithms: logistic regression, support vector machine (SVM), random forest, K neighbours and decision tree. Meanwhile, in the second approach, used an artificial neural network with intensive use of network optimization and hyperparameter tuning to achieve the highest accuracy possible. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Detection of abnormality in Musculoskel et al. radiographs often requires the involvement of a radiologist but the recent advancement in deep learning technique helped us to detect abnormality in musculoskeletal radiographs. In this paper, each model’s performance is evaluated on unseen Mura-dataset. A novel method has been developed based on CNN architectures (DenseNet169, Vgg16 and ResNet50) by training it with MURA dataset and performed hyperparameter optimization for all the 3 models. Showing how Deep Convolution Neural Networks could be extended to medical images. DenseNet169 showed the highest accuracy. In this paper, found that model (Densenet169) achieved training accuracy of 87.88% and validation accuracy of 79.20%. In case of training loss and validation loss, the model achieved 0.3304 and 0.4887 respectively.

Self driving cars are picking up pace and thus serves as a challenging research domain. One of the most crucial functions in an autonomous vehicle is accurately detecting and recognising the traffic signs. Traffic sign detection is a crucial task in traffic sign recognition systems. Deep neural networks are proven powerful in traffic sign classification. As traffic sign violation leads to law and order disruption, posing a threat to human life, there is a need for robust algorithms with efficient actuation to perform the delegated task. Thus, this paper proposes a robust algorithm which addresses challenges like haze, fog, unclear images due to improper condition of roads to efficiently detect and recognize traffic signs using Image manipulation, Optical Character Recognition (OCR) algorithm and You Only Look Once (YOLOv3) detection algorithm. The proposed algorithm in this paper for hazy images that aids in ADAS applications has an accuracy of 87.29%. This contributes to the emerging research domains of autonomous vehicles and self-driving cars.

Background. Online platforms have now fully emerged as a platform to share their views on various subjects and issues with various viewers. Purpose. Though we have different ways of expressing ourselves, the most common is text. The dataset has texts, selected_text, punctuations, emoticons, etc. Methods. The paper targets to analyze sentiments from Twitter posts. We use the approach to train our model for detecting sentiment from tweets. Some interesting novelties in this paper are (i) including a hybrid approach for analyzing the sentiment of the tweets, (ii) introducing Jaccard similarity, and (iii) aiming to find which neural network (RNN or CNN) algorithm works better with our dataset. Results. With overall analysis, Bi-Directional LSTM performs better out of the three, and hence, the accuracy of the hybrid model enhances, indirectly. Conclusion. Finally, a concluding scope in which RNN and our hybrid approach perform better and are approximately 4–5% improved than the past models. The future implications related to the work are also given to improve the quality of the model.

The paper deals with the issues of automatic assessment of the quality of glass rods in the production process. Traditionally, quality control at glass enterprises is carried out visually by a process control operator. The high speed of modern technological processes leads to errors in quality assessment. To improve the quality of glass rods, an automatic control system based on machine vision methods is required. Such a system should automatically recognize defects in digital images that continuously come during the production process. The main task of image processing in the quality control system is the automatic search and recognition of defects. The proposed solution to this problem includes two stages. At the first stage, a set of features (descriptors) is formed, which are necessary for the classification of defects. The paper describes the presentation of features based on the use of statistical methods of segmentation of textured images. For the formation of descriptors, the information contained in the histograms of the first and second orders was used. A grayscale difference matrix was also used. The procedure for validating and classifying defects was carried out using a deep learning neural network. The description of the network and the algorithm for its training are given. In order to assess the effectiveness of the automatic defect recognition system, tests were carried out. The results of these tests showed that the use of a neural network and the proposed representation of features provides good discrimination for the analysis and classification of glass rod defects.

Facial Expression recognition (FER) is a vital field of artificial intelligence and computer vision owing to its remarkable academic potential. Recognition of facial expression has many implicit applications, and it has attracted much attention of researchers during the last decade. The relation between the facial features to each human emotion is the main focus of the investigation. By training a neural network with such features, one can realize a system that can provide analytical data on human behaviors. This paper proposes a facial expression recognition system architecture based on a deep convolutional neural network. In the course, various neural networks are studied and compared on their facial expression recognition, and one can find that the CNN deep learning method can be one of the best options to take. The paper includes human facial emotion detection procedures that include three significant steps: face recognition; characteristic feature extraction; emotion classification. A set of experiments is performed to evaluate the proposed system’s performance, and the recognition accuracy is measured in terms of the facial emotion recognition challenge (FERC-2013) and Japanese female facial expression (JAFFE) dataset.

Only 4% of household waste generated in Africa is recycled. Current research uses machine learning models in cloud-based solutions to classify waste. However, in countries with limited internet access, there is a need to increase user engagement in classifying waste using an on-device approach. Developing a machine learning model for a mobile device with limited size and speed is a challenge. This research proposes an on-device deep learning framework to encourage the recycling of household waste. The proposed framework combines an optimal deep learning image classification model and gamification elements. A combination of multiple waste datasets named WasteNet consisting of 33,520 images is used to train the deep learning image classification model using seven classes of recyclable waste namely e-waste, garbage, glass, metal, organic, paper and plastic. Data augmentation and transfer learning techniques are applied to train five models on a mobile device namely, MobileNetV2, VGG19, DenseNet201, ResNet152V2 and InceptionResNetV2. Results of the five models are presented in this paper based on accuracy, loss, latency and size. This research shows promise for InceptionResNetV2, MobileNetV2 and DenseNet201in encouraging householders to engage in recycling waste using gamification on a mobile device.

Advanced driver-assistance systems (ADAS) play a significant role in reducing traffic fatalities and vehicular accidents. Among the many ADAS technologies is the blind spot alert monitor alerts you of a vehicle or an obstruction before you change lanes. This paper investigates extending the monitoring concept to detect if a car emerging from the blind spot is attempting to cut into your lane. In that respect, we would perceive the direction of the infringing vehicle. There are various means of sensing such a situation, including radar, lidar, camera, and sonar. We focus on using lidar and deep learning for this purpose in this paper. In particular, we present the experiments carried out to verify the concept, including details of the equipment, experimental data, and MATLAB Deep Learning for detection.

Roads make a huge contribution to the economy of a territory. As a platform for transportation, roads are widely used by every countries in the world. Potholes in road are one of the major concerns in the transportation infrastructure. A lot of research works have been proposed using Computer Vision to detect potholes that include wide range of image processing and object detection algorithms. There is a need for potholes detection with adequate accuracy and speed, and that can be implemented with ease and low setup cost. In this paper, we have developed efficient deep learning Convolution Neural Networks (CNNs) to detect potholes in real-time with adequate accuracy. This paper compares the performance of YOLOv5 Large (Y\(_{l}\)), YOLOv5 Medium (Y\(_{m}\)) and YOLOv5 Small (Y\(_{s}\)) with ResNet101 backbone and Faster R-CNN with ResNet50 (FPN), VGG16 and MobileNetV2 backbone. The experiments results show that YOLOv5s is more applicable for real-time potholes detection because of its speed. It is able to detect potholes in high as well as low resolution images in 0.009s. With the use of Inverse Perspective Mapping with YOLOv5 to estimate area, we concluded that the detection can still be done in real time with area estimation.

Fine-tuning the pre-trained language model is the current mainstream method of text classification. Take the fine-tuning BERT model as an example, this kind of approach has three main problems: the first one is that the training of massive parameters will cause high training costs The second is that the model is very easy to over-fit in trainable samples, resulting in low transferability. Third, the fine-tuning model is not good at long text classification task. In this paper, we take the sentiment classification task as an example and use the classification accuracy as a metric. We compared two methods to problem one: using a compressed language model (decreased 0.1%) and using entirely frozen weight (reduced by 4%). For the second problem, in the case of using fixed weights (reduced by 4%), Convolutional Networks (CNN) and Capsule Networks (CAP) are used to extract n-gram features and clustering features so that the classification accuracy is improved (increased by 0.5% and decreased by 0.2% respectively). The corpus transfer test shows that CAP’s accuracy is greater than the fine-tuning model (increased by 0.6%). Meanwhile, this article proposes a method to process the long text classification task by expanding position embedding to support long text input (increased by 1.1%). Finally, this paper compares the training speed and parameter scale of different models under the combination of different strategies, and uses the F1 value, precision rate, recall rate and accuracy to measure each model.

This paper proposes a standard way to represent sparse tensors. A broad theoretical framework for tensor data scattering methods used in various deep learning frameworks is established. This paper presents a theorem that is very important for performance analysis and accelerator optimization for implementing data scattering. The theorem shows how the impossibility of slicing happens in tenser data scattering. A sparsity measuring formula is provided, which can effectively indicate the storage efficiency of sparse tensor and the possibility of parallelly using it. The source code, including CUDA code, is provided in a related open-source project.

The paper presents a new approach to recognize small parts from high-resolution 4K images grabbed by the embedded camera with trained convolutional network models. The main principle is a combination of standard machine vision techniques (OpenCV library) to identify blobs in the high-resolution image and parse to a suitable set of regions of interest, which are later recognized by trained convolutional neural network model (Tensorflow framework). The convolutional neural network model with single-shot detection was selected to identify the position of the object. 3D virtual models of parts with autogenerated orientations are used for simpler input data generation. Two embedded devices were selected for use in the experimental system. The first is NVidia Xavier dev board with 4K camera and integrated APU and the second is Raspberry PI 4 with external APU Intel Neural Compute Stick 2. The last part of the paper shows the implementation of the presented principle to the virtual device (HTC Vive Pro) for assisted assembly tasks to train and check workers during the teaching process.

As on 15th September 2020, the total cases of SARS-CoV-2 infected patients in the world has crossed 29 million, with more than 930,000 deaths occurring due to the virus. Real time RT-PCR (Reverse Transcription Polymerase Chain Reaction), which is the standard detection method for COVID-19, is likely to have low detection rate in early stages of the infection possibly due to less viral load in the patients. On the other hand, in comparison to RT-PCR, patterns obtained from radiography on chest CT scans show higher sensitivity and specificity. However, due to the sensitive nature and difficulties in publicly acquiring medical data, only 2 open-sourced COVID CT datasets with images containing clinical findings of COVID-19 could be found. Applying existing deep learning models to the limited CT scans can distinguish COVID-19 and non-COVID-19 CT images but with lesser accuracy. The paper proposes to use existing deep learning-based frameworks on an augmented dataset consisting of pixel scaled images (of the original CT images) and the original CT images to diagnose COVID-19 infection. Since this is a binary classification problem the paper proposes to use Convolutional Neural Networks (CNN) to classify CT images into infected and not infected categories. The implementation is done with Keras and Tensorflow using an 80/20 ratio for training and validation. The proposed methodology (using pixel scaled images) achieved a validation accuracy of over 90% in detecting COVID-19 with an F1-score of 0.96 compared to the best F1-score of 0.86 on the original dataset.

To manage their disease, diabetic patients need to control the blood glucose level (BGL) by monitoring it and predicting its future values. This allows to avoid high or low BGL by taking recommended actions in advance. In this paper, we conduct a comparative study of two emerging deep learning techniques: Long-Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) for one-step and multi-steps-ahead forecasting of the BGL based on Continuous Glucose Monitoring (CGM) data. The objectives are twofold: 1) Determining the best strategies of multi-steps-ahead forecasting (MSF) to fit the CNN and LSTM models respectively, and 2) Comparing the performances of the CNN and LSTM models for one-step and multi-steps prediction. Toward these objectives, we firstly conducted series of experiments of a CNN model through parameters selection to determine its best configuration. The LSTM model we used in the present study was developed and evaluated in an earlier work. Thereafter, five MSF strategies were developed and evaluated for the CNN and LSTM models using the Root-Mean-Square Error (RMSE) with an horizon of 30 min. To statistically assess the differences between the performances of CNN and LSTM models, we used the Wilcoxon statistical test. The results showed that: 1) no MSF strategy outperformed the others for both CNN and LSTM models, and 2) the proposed CNN model significantly outperformed the LSTM model for both one-step and multi-steps prediction.