International Symposium on Intelligent Informatics

We present a novel architecture for classifying imagined words from electroencephalogram (EEG) captured during speech imagery. The proposed architecture employs a sliding window with overlap for data augmentation (DA) and common spatial pattern (CSP) in order to derive the features. The dimensionality of features is reduced using linear discriminant analysis (LDA). Long short-term memory (LSTM) along with majority voting is used as the classifier. We call the proposed architecture the DCLL (DA-CSP-LDA-LSTM) architecture. On a publicly available imagined word dataset, the DCLL architecture achieves an accuracy of 85.2% for classifying the imagined words “in” and “cooperate”. Although this is around 7% less than the best result in the literature on this dataset, the DCLL architecture is roughly 300 times faster than the latter, making it a potential candidate for imagined word-based online BCI systems where the EEG signal needs to be classified in real time.

Recently, it has been stipulated that training larger and larger models, using ever increasing datasets is not sustainable in a long-run. Hence, the idea of Frugal Artificial Intelligence has been put forward. While there are many ways to make AI frugal, this contribution focuses on two of them, namely neural network pruning and binarization. Experimental results obtained for the CIFAR-100 and CalTech-256 datasets are presented and analysed. Several issues related to practical implementation of pruning and binarization, stemming from the performed experiments, are pointed out.

Despite the promise for performance and accuracy improvements of quantum inspired (QI) algorithms over classical machine learning (ML) algorithms, such gains have not been realized in practice. The quantum inspired algorithms can theoretically achieve significant speed up based on sampling assumptions and have thus far failed to outperform the existing classical ML models in practical applications. The speedup of quantum machine learning (QML) algorithms assume the access to data in quantum random access memory (QRAM) which is a strong assumption with current quantum architectures. QI algorithms assume sample and query (SQ) access to input vector and norms of matrices using a dynamic data structure. We explore the components of these models and the assumptions in this paper by surveying the recent works in QML and QI Machine learning (QIML) algorithms. We limit our study to QML and QIML models on achieving a speed up over classical ML techniques rather than individual proofs of these algorithms. This study highlights the assumptions being made that are currently not practical for QML and QIML algorithms in achieving performance advantage over classical ML algorithms.

The robustness and applicability of the Encoder–Decoder Network with Guided Transmission Map (EDN-GTM) proposed for efficient single image dehazing purpose are examined in this paper. The EDN-GTM utilizes the transmission map extracted by dark channel prior approach as an additional input channel of a novel U-Net-based generative network to achieve an improved dehazing performance. The EDN-GTM has shown a very favorable performance compared with most recently proposed dehazing schemes including both traditional and deep learning-based ones in terms of PSNR and SSIM metrics. To further validate the robustness and applicability of the EDN-GTM scheme, extensive experiments and quantitative evaluations on various benchmark datasets are conducted in this paper. In terms of robustness, experimental results on different benchmark dehazing datasets such as Dense-HAZE, NH-HAZE, and D-HAZY show that the EDN-GTM scheme consistently outperforms most modern dehazing approaches on both synthetic and realistic hazy data regardless of scene locations: indoor or outdoor. On the other hand, experiments on WAYMO and Foggy Driving datasets imply that the EDN-GTM can be effectively applied as an image pre-processing tool to object detection tasks in autonomous driving systems.

Rotating machines have components like bearings, rotors, stators, etc. Among these components, bearings are the most critical component and also the primary source of failure in any machine. So, monitoring the health of bearings becomes important to prevent any catastrophic machine failures and avoid unpredicted machine downtime. This paper uses open-source data from the Case Western Reserve University (CWRU) to monitor the health of bearings. For this work, the time-domain features are extracted on raw, differentiated and integrated signals followed by different fusion techniques before passing through a neural network. One such technique is decision fusion, in which all the three different neural network outputs for fault classification are fused for bearing fault detection. Also, a Graphical user interface (GUI) and real-time simulation have been developed for ball-bearing fault diagnosis using MatLab.

Nowadays, the smallest of things can spread like wildfire. The hashtags highlight and provide an easy way to gain knowledge on a particular topic. It is used to help a person in getting all the data there exists on that particular platform about the topic they are looking for. This article gives an overview of the various fashion production movements prevalent on social media that helps in spreading awareness to a wider audience. Based on the online data analytics (web app and mobile app data), the article aims to correlate social media, fashion, and the environmental impact of these movements. When it comes to the production of fashion goods, many people might not be aware of the issues related to it. Climate change, animal rights, and worker’s rights are topics along with their movements which have been discussed in this article. The fashion industry is the second largest polluting industry in the world. Most of the production of garments takes place in third world countries such as China, Bangladesh, Indonesia, Vietnam, and India. Through this research, an attempt has been made to fill the gap between consumers’ awareness, their positive attitude toward sustainability and exemplary fashion, and a lack of action while purchasing decisions.

Large Scale Global Optimization (LSGO) is an identified problem in the literature and almost all Bio-inspired metaheuristic search-based optimization algorithms, including Genetic Algorithm (GA), face this problem. The article first identifies schema deception, domino convergence with genetic drift and nonseparability among the variables as the reasons for LSGO. Towards LSGO solution, the article progresses the cooperative coevolution approach, with some novel concepts in solution representation, subcomponent selection, search mechanism and static adaptation of algorithm parameters. The concepts are demonstrated on GA as it is having a strong theoretical and mathematical background. Many GA variants are derived based on the above novel concepts those better balance the exploration and exploitation abilities of an algorithm, specifically at large scale. The proposal is justified using a performance comparison with other algorithms on some of the LSGO test bench functions with 4–20 dimensions.

Chaotic systems are sensitive to initial conditions. The exponential divergence of nearby trajectories limits the accuracy of the long-term prediction of a chaotic time series. Because of the deterministic governing equations of the underlying system, accurate short-term prediction of a chaotic time series is possible. Various approaches have been used to forecast a chaotic time series, the most popular of which is the first-order approximation of the local dynamics in the embedded state space. Simultaneously, various strategies for modeling a time series by a complex network have also been developed. Nonetheless, time-series induced networks have received little attention in terms of forecasting time series. This paper proposes a method based on symbolic dynamics for constructing a weighted network from a given time series and provides a strategy for forecasting a time series using the weighted network. We demonstrate the approach’s effectiveness by predicting a chaotic time series. The results are then compared to those obtained using the linear first-order approximation method. The proposed method is straightforward, computationally efficient, and parameter-free.

In computational neuroscience, the automatic fitting of neuron parameters of simple neurons based on data or other detailed models is a challenging problem. Stochastic variability of firing activity in neurons makes an automated fitting, a complex neuronal dynamic problem. In this study, four features of neuronal firing including spike amplitude, spike count, spike timing, and inter-spike interval were employed to create a simplified neuronal network model using particle swarm optimization and genetic algorithm. This paper showcases a methodology to reduce detailed models that arise from experimental data into simple spiking Izhikevich models, which can be computationally more effective while reconstructing large-scale circuits for behavioral modeling. The approach is illustrated during spontaneous firing activity and current injection to validate the model. The result indicated that the reduced neuronal model shows matching firing activity when optimize with optimization algorithms. The four features of the neuronal activity were matched with the experimental data after optimization. The study also analyzed the parameter fitting accuracy and runtime efficiency of reduction based on the optimization algorithms. The result indicated that the reduction based on particle swarm optimization showed less error percentage while reducing the model when compared with the genetic algorithm.

Handwriting digit recognition is a computer technology that allows it to accept and decipher sound transcribed input from various sources, including paper reports, contact screens, and pictures. This paper presents the machine and Deep Learning approach for handwritten digit recognition from image input. This approach uses bench marked dataset MNIST English handwritten numeral digit dataset of size 70,000. Four machine learning and deep learning algorithms are explored, and a pattern recognition approach is used. We have explored the pattern matching approach and achieved 86% accuracy for the Decision Tree classifier, 91% accuracy for the Support Vector Machine classifier, 97% for Artificial Neural Network and 98.84% for Convolutional Neural Network. In the Deep Learning approach, the Convolutional Neural Network algorithm with the Vgg16 network is implemented to train the MNIST digit dataset and achieved an accuracy of 98.84%.

Malicious disease cancer is predominant and spreading like any other epidemic disease. Awareness on preventive and post-preventive measures is inevitable to combat this syndrome. Presently, bone cancer exceeds its limit and the gravity of attention must be focused on digital remedy. This research work is framed with major bone cancer namely Ewing sarcoma. In this direction, the basic features of Ewing sarcoma are studied, and the algorithm for automated diagnosis is developed. The algorithms (1) model-based clustering and (2) gradient-based watershed segmentation are developed to extract the nucleus and cytoplasm. To ensure quality segmentation, H &E components are separated by the deconvolution method. Parallel application of segmentation algorithms is used on the H &E component for better time factor. The features are defined to construct a classification model by using a Support Vector Machine. There are 300 images of different features selected to train the SVM classifier and 120 images are used for testing and the accuracy obtained is 91.6%.

Deep Neuroevolution (DNE) holds the promise of providing radiology artificial intelligence (AI) that performs well with small neural networks and small training sets. We seek to realize this potential via a proof-of-principle application to MRI brain sequence classification. We analyzed a training set of 20 patients, each with four sequences (weightings): T1, T1 post-contrast, T2, and T2-FLAIR. We trained the parameters of a relatively small convolutional neural network (CNN) as follows: first, we randomly mutated the CNN weights. We then measured the CNN training set accuracy, using the latter as the fitness evaluation metric. The fittest “child” CNNs were identified. We incorporated their mutations into the “parent” CNN. This selectively mutated parent became the next generation’s parent CNN. We repeated this process for approximately 50,000 generations. DNE achieved monotonic convergence to 100% training set accuracy. DNE also converged monotonically to 100% testing set accuracy. DNE can achieve perfect accuracy with small training sets and small CNNs. Particularly when combined with Deep Reinforcement Learning, DNE may provide a path forward in the quest to make radiology AI more human-like in its ability to learn. DNE may very well turn out to be a key component of the much-anticipated “meta-learning” regime of radiology AI; algorithms that can adapt to new tasks and new image types, similar to human radiologists.

Text Summarization based on Abstraction is the task of generating a concise summary that captures the principal ideas of the source text. It potentially contains new phrases that do not appear in the original text. Although it is widely studied for languages like English and French, owing to the scarcity of data on regional vernacular languages like Hindi, the research in this area is still in the primitive stages. We propose a novel approach for building an Abstractive Text Summarizer for Hindi corpus using the Transformer encoder-decoder architecture. Firstly, efficient pre-trained word representations are generated using Facebook’s fastText model. Next, the Transformer model is employed to extract contextual dependencies and yield better semantic representations for a morphologically rich language like Hindi, engendering an abstractive summary. On performing an experimental evaluation on the Hindi news dataset to generate news article headlines, we achieve a ROUGE-1 precision and recall score of 0.682 and 0.598, respectively, which outperforms the state-of-the-art techniques.

Text documents are important sources of information that offer insights to administrators, planners, and researchers on what the public likes and does not. They are available in the form of plan documents, government reports, journal write-ups on sports, politics, science, and technology, etc., at web portals. Since its quantity is huge, searching for them on the web is time-consuming. This calls for a real-time tool that would be quicker and effective with categorizations. Comparisons with legacy documents that provide public opinions are constant or changing over a period of time. This study examines existing processes involved in article categorization and puts forth new algorithmic techniques for the automatic classification of Malayalam web-based articles. The performance of new classifiers is evaluated by using precision, recall, accuracy, and F1 score values.

Natural language generation is one of the major tasks that comes under Natural Language Processing and, Question Generation (QG) and Question Answering (QA) are two NLP-based tasks that have found widespread application. The QA and QG tasks on a text paragraph aim to extract answers and generate questions, respectively. The automatic answering systems, chatbots, and trivia applications make use of these tasks and their use cases are growing. We have used a text-to-text transformer, T5 with a transfer-learning-based approach to achieve these tasks by training a single model in a multi-task fashion. The study is based on fine-tuning a T5 transformer in a multi-task way to produce questions from the text and extract answers upon providing the questions. The fine-tuning step is used to train the pre-trained transformer on data-science-related text paragraphs along with a large question answering dataset. This work aims to implement simplified data processing and training for fine-tuning the transformer model to improve the performance of the pre-trained model on technical data for QA and QG tasks.

In recent years, with technical improvements and growth in content-based advertising, individuals have started utilizing SMS (Short Message Service), which has resulted in a significant increase in spam SMS. African continent experiences the most spam SMS in the globe, with an average of 119 spam SMS received per month by a person. These spam SMS are unsolicited messages to users, which are disturbing and may sometimes lead to the loss of important data. There exist many spam SMS detection methods which are impacted by the inclusion of well-known words, phrases, abbreviations, and idioms. The proposed work is using supervised machine learning techniques such as Multinomial Naïve Bayes and Support Vector Machine to identify SMS as spam or ham and compare their results based on specific evaluation parameters to find the most effective technique for filtering messages. The technique is evaluated on a real-world SMS dataset including over 5572 messages. According to the findings, the Support Vector Machine algorithm is the best at classifying SMS as spam or ham with an accuracy of 98.83%.

With the changing behavior of different types of social networking sites, such as Instagram, Twitter, SnapChat, etc. the amount of data shared by people, or users of a particular social site, is rapidly expanding. Every day, around million and billion pieces of data are uploaded. This is because a certain website has millions of users. These folks want to offer their thoughts and opinions on any topic they want. We hope to deduce the feelings behind these posts in this study. It is difficult to collect and analyze people’s reactions to purchasing a product, using public services, and so on. Sentiment analysis is a frequent debate preparation work that seeks to uncover the sentiments that underpin opinions in various texts. In recent years, sentiment analysis researchers have focused on assessing opinions on a variety of topics, including movies, commercial products, and everyday societal challenges. Twitter is one of the most popular micro-blogs where customers may express themselves. Opinion research using Twitter data has gotten a lot of attention in the previous decade. The two methodologies for assessing feelings from the text are first one is the knowledge base approach, and the second one is machine learning based approach. In this work, we use a Machine Learning based technique to evaluate tweets on electrical devices, such as smartphones and laptops. It is feasible to determine the influence of domain information on sentiment categorization by performing sentiment analysis in each domain. A novel feature vector for categorizing of tweets as positive or negative, as well as extracting people’s opinions on items has shown. Paper presents an analysis of popular methodologies for opinion mining, such as machine learning and lexicon based approaches, as well as assessment measures.

Blood Vessels have a significant role in the diagnosis of Diabetic Retinopathy (DR) through retinal images. However, the major issues are the accurate segmentation of blood vascular structure from the retinal image. As there exist tiny vessels in the retina at the advanced stages of DR, the extraction of such kind of vessels is a challenging task. Hence, this paper proposes a new retinal vasculature segmentation mechanism based on pixel-wise classification. A new feature vector called as Cascaded Feature Vector (CFV) is introduced here to represent each pixel with a set of composite features. To extract such features, this approach totally employs five different filters namely Edge (E), Morphology (M), Statistical (S), Hessian (H), and Gradient (G) filters. Based on obtained features, CFV is formulated and fed to machine learning algorithms for classification. Artificial Neural Networks (ANN) and Support Vector Machine (SVM) are employed for classification. Experimental validation on the two datasets namely DRIVE, and ARIA proves the effectiveness of the proposed method in terms of segmentation accuracy.

Lesion segmentation in medical imaging is key to evaluating treatment response. We have recently shown that reinforcement learning can be applied to radiological images for lesion localization. Furthermore, we demonstrated that reinforcement learning addresses important limitations of supervised deep learning, namely, it can eliminate the requirement for large amounts of annotated training data and can provide valuable intuition lacking in supervised approaches. However, we did not address the fundamental task of lesion/structure-of-interest segmentation. Here we introduce a method combining unsupervised deep learning clustering with reinforcement learning to segment brain lesions on MRI. We initially clustered images using unsupervised deep learning clustering to generate candidate lesion masks for each MRI image. The user then selected the best mask for each of 10 training images. We then trained a reinforcement learning algorithm to select the masks. We tested the corresponding trained deep Q network on a separate testing set of 10 images. For comparison, we also trained and tested a U-net supervised deep learning network on the same set of training/testing images. Whereas the supervised approach quickly overfits the training data and predictably performed poorly on the testing set (16% average Dice score), the unsupervised deep clustering and reinforcement learning achieved an average Dice score of 83%. We have demonstrated a proof-of-principle application of unsupervised deep clustering and reinforcement learning to segment brain tumors. The approach represents human-allied AI that requires minimal input from the radiologist without the need for hand-traced annotation.

Emotion recognition based on biological signals has been a hot research topic and it has many practical applications in various areas such as safe driving, health care, and social security. This work aims to examine the performance of the Random Forest Classifier, Logistic Regression, SVM, and XGBoost Classifier model for EEG-Based Emotion Recognition. To measure the performance of these models, we used the Emotions dataset, which is publicly available and consists of pre-processed physiological signals. Among the proposed four models, the XGBoost Classifier achieves the best classification accuracy of 99.12%, which indicates a significant increase in the classifier’s performance applied to the dataset.

Machine vision applications for intelligent vision systems in manufacturing industries were reported based on image processing and artificial intelligence technology. We propose the imaging and vision development platform in this research for creating vision applications using image processing, machine learning, and a deep learning algorithm library. An algorithm library, vision configurator, execution logic, display manager and deploy manager modules are all included in the proposed platform. This platform is based on an open-source software stack for machine learning and deep learning computer vision technologies including OpenCV, TensorFlow, CUDA, Keras, YOLO and PyTorch. To assess the performance of the suggested platform, real-time applications like vehicle identification, person detection, code scanner, and OCR vision application were developed, validated, and deployed in an embedded system utilizing this platform. The results of the experiments show that the suggested platform can be utilized to evaluate high resolution real-time images and construct vision applications.

Diabetes is a biochemical dysfunction that disrupts the productivity of work and can severely affect the way of living life through bounded and strict diet charts. Blood Glucose levels can be detected via many physiological variables like measurement of insulin and pulse waves among others. In this paper, photoplethysmography (PPG) signal and its features using pulse decomposition analysis (PDA) for developing a non-invasive diabetes detection system is studied. For this study, dataset of 217 participants from normal, pre-diabetic and diabetic individuals is used. In addition, three physiological parameters are also used i.e., age, weight and height for the classification. Database is filtered and segmented into 5 s segments. Support vector machine (SVM) classifiers and neural network (NN) classifiers are used for the classification of the dataset into three classes that is diabetic, pre-diabetic and non-diabetic. A good accuracy is obtained with NN classifiers and the study shows that for the non-invasive blood glucose detection, PPG signal from wearable sensors can be an acceptable option.

Noise removal is mandatory to attain better clarity in images and videos, especially in surveillance, military, and forensic applications. This research proposes a noise classification algorithm using statistical features of the noise added to the videos. The classified noise is then eliminated with the help of a suitable filter and then recovered using a novel, non-iterative pseudo inverse-based recovery algorithm (NIPIRA), a compressive sensing-based recovery algorithm. The statistical parameters namely, total mean (TM), total variance (TV), total standard deviation (TSTD), and skewness (skew) are used as noise-classifying parameters based on which the appropriate filter is chosen. The recovery rate or precision in recovery using the classification and filtering process is as high as 99% while using NIPIRA. The PSNR obtained after recovery and filtering is approximately 36 dB, which is about 10 dB higher than the existing compressive/compressed sensing (CS) based noise removal algorithms such as orthogonal matching pursuit with cross-validation (OMP-CV). The structural similarity between the input and reconstructed images and videos was preserved better by NIPIRA than by the existing algorithms. Hence the proposed method is much suited for noise removal and fast reconstruction of surveillance videos.

The security of digitally shared data like images, audio and video incurs many security and privacy-related challenges such as unauthorized access. In this work, a video encryption technique using improved sine map and DNA sequencing has been proposed. The proposed video encryption process mainly consists of three phases. In the first phase, key is generated using SHA-256. The second phase comprises of permutation of the video frame using SHA-256 key and improved sine map. In the final phase, diffusion is performed of the permutated frame by using SHA-256 and DNA sequencing, to obtain encrypted video. The experimental outcomes unveil that the obtained entropy of the encrypted video frames is approximately 8 that reflects high randomness in the encrypted frame, and it has low correlation coefficients among neighbouring pixels in all three horizontal, vertical and diagonal directions. The presented scheme of video encryption is resistant to threats like brute force attack, plaintext attack and statistical attack.

Medical image processing is one of the significant fields to identify the diseases as earlier to diagnose them appropriately. The brain tumor segmentation process is sub branch of a medical image processing field. The computer vision and machine learning techniques provide an effective channel for the medical practitioners for diagnosing the diseases in an effective method. This research article implements the Srgb-based density analysis for isolating the brain tumor space in MRI images. Intensity values of a given input are normalized using Srgb color space and Gaussian filter to distinguish the tumor region from the background. The adaptive threshold technique helps identify the possible tumor space in brain MRI samples. The actual brain tumor space is extracted by performing the region properties such as area and density function. Finally, the accurate tumor space is detected by applying morphological functions with eliminating possible false positives. Performance metrics including recall, precision, and F-measure are used to assess the effectiveness of the proposed approach.

Lung nodule segmentation is an essential aspect of two separate systems which deal with the diagnosis and prevention of lesions. A better computer-aided diagnostic (CAD) system attempts to increase the capacity to detect nodules and assist in determining whether they are malignant or benign. Lung segmentation precisely distinguish the areas and borders of the lung field from neighboring thoracic tissue and is indeed an important initial step in several clinical decision support methods. In pulmonary image analysis, one of the essential tasks is to segment the lung nodules that aid as an essential step for clinicians in distinguishing malignant and benign tumors. Yet, it is being noted that this might be a challenging task, particularly while nodules are attached to anatomic structures. This research presents a novel lung nodule segmentation method, which is executed in two steps: filtering and segmentation. The input lung image is pre-processed by employing an improved Wiener filtering and this filtered image is subjected to segmentation using an improved Kapur entropy-based segmentation method. The segmented output is then acquired in a practical manner. The outcome of the developed technique is compared to existing techniques in terms of various metrics.

For efficient storage and transmitting of medical images over the internet, an adequate compression method is necessary. Modern healthcare services that rely on cloud computing use such compression methods, as often we are limited by internet bandwidth and memory to store these medical images on cloud servers. The Digital Imaging and Communications in Medicine (DICOM), which is the standard currently used for transmission and handling communication of medical data uses the JPEG-2000 standard for compression. Recent impressive advancement in computer hardware and software leads to new compression methods to be developed and used widely for many applications. High-Efficiency Video Coding (HEVC/H.265/x265) is a video coding standard that has intra-frame coding for still medical images such as X-rays, MRIs, CT Scans, etc and as an advanced video coding standard, it supports compression of medical videos such as ultrasound and sonography. HEVC can be a single format that can support the compression of still Medical images, images series, and videos as well as modern medical imaging such as 3-D X-rays and 3-D Ultrasound. This study illustrates the utilization of HEVC for the compression of Medical images and a comparison of various compression techniques that have been used here for medical image compression. The Lossless compression performance of HEVC is compared to JPEG -2000 for grayscale and RGB images. Experimental results show, that in lossless mode, HEVC performs up to $${\sim }15\%$$ better for grayscale and $${\sim }17\%$$ better for RGB images. For lossy mode, HEVC accomplishes better results than JPEG, JPEG-2000 and AVC.

In high efficiency video coding intra coding framework is built on the prediction of spatial sampling. Intra prediction is used in high efficiency video encoder to remove spatial redundancy. The intra prediction operates according to transform block size. The previously encoded boundary sample of the neighboring transform block is used as the reference pixel. The spatial redundancy is removed by predicting the pixels in the region of the coding block from the reference pixel. Copying-based method is used for intra prediction process in high efficiency video encoder (H.265). The proposed research work is carried out in two phases. In the first phase theoretical analysis of intra prediction block is performed with due consideration of the first-order Gaussian Markov model as a reference under two distinct conditions: 1. The pixel values differ from the model values to large extent and 2. The distance between the reference pixel and predicted pixel is too large. Both of these cases are evaluated by finding the correlation between reference pixel and coding block pixel. The coding gain is used as the performance parameter which indicates the prediction accuracy of the intra prediction process. The analysis shows that in both of these conditions the corresponding correlation is very weak and the coding gain declines, in such conditions prediction weight should be very small. In the second phase paper proposed an improvement to copying-based method in the second phase. The repetitive smoothing filter is added as an extra mode to the traditional copying-based method. The number of iterations of the filter is adaptively changed based on the correlation between the reference and block pixels. The repetitive filter with varying number of iterations number reduces the number of encoding bits up to 2.5% on high resolution video sequences. In the repetitive filtering number of times the filter is iterated is decided by the correlation between current and reference pixel. The significant reduction in the number of bits increases the coding gain of the high efficiency video encoder than the default value of the coding gain.

Doctors view the total number of leukocytes in a person’s blood as a crucial sign of that individual’s general health. Historically, blood cell counting was performed manually using a hemocytometer and a few more lab equipment and chemicals. The process is slow and laborious. Red blood cells (RBC), white blood cells (WBC), and platelets are the three kinds of blood cells that can be detected and counted automatically using picture segmentation and S-CNN (Suit-Convolutional Neural Network) machine learning algorithms, as demonstrated in this body of work. Red blood cells, white blood cells, and platelets may be located and counted automatically using an open-source database of blood smear pictures. When the trained model was evaluated using smear images from a distinct dataset, it was observed that the learned models were relatively simplistic. The computer-aided tracking and identification technique allows us to count blood cells in less than one second from photographs of character assassination. It is helpful for real-world applications and has an approximate 92% degree of accuracy.

Artificial intelligence is making a huge impact in terms of accuracy, robustness, and real-time capabilities in advanced driver assist systems. The conventional RaDaR systems using routine signal processing (SP) techniques are not able to meet such high requirements when it comes to the classification of the target’s intent. Here a strategy that combines SP and machine learning (ML) techniques to detect dynamic changes in target morphologies with the help of micro-Doppler (MD) signatures is proposed. The target’s MD signatures are used to create target libraries, which will be further used to train the ML model for accurate classification of the target’s intent. Finally, the overall raw data generated is implemented on an edge computing platform aimed at a future deployment of the ML model in field-deployed computing devices.

Since knowing human emotion is one step closer to artificial intelligence, more possibilities for emotion detection are being explored. Emotion recognition is a difficult task, and it becomes even more challenging while using speech. Though humans can interpret emotion from facial expressions, body language, and speech patterns, it is difficult for machines to do so. We have suggested a technique for detecting emotion from voice signals in this research paper. We looked at three separate datasets in this study. For our research, we use numerous audio features such as MFCC, Chroma, ZCR, RMS, and Glottal to determine which one will be more useful. The efficiency of various machine learning algorithms and neural network trained using audio features was calculated and trained at different parameters and hyper-parameters. Using the confusion matrix, the accuracy of various methods was compared. Then we matched our findings to those of previous research. The majority of the emotions were also appropriately classified. Then we attempt to create a system that is language agnostic.

A cornerstone of wireless connectivity involves trust and privacy in the data shared between users and network elements as wireless connectivity becomes an integrated, fundamental element of society. With a large influx of data in Beyond 5G (B5G) systems from end-users and network elements, it is imperative to understand how data is collected and used for real-time data processing operations. The current wireless network learning involves centralizing the training data, which is inefficient as it continuously requires end devices to send their collected data to a central server. Federated Learning (FL) effectively allows end devices to train ground-truth data on-device, and only model update parameters are sent back to the federated server. This work proposes a Chameleon FL model, FED6G, for network slicing in 5G and Beyond systems to solve complex resource optimization problems without collecting sensitive, confidential information from end devices. The evaluation results reflect more than 39% improvement in Mean Squared Error (MSE), 46% better model accuracy, and more than 23% reduced energy cost for training the proposed FED6G against the traditional deep learning neural network model.

Recently, clustering is emerged as an optimal solution to improve the network lifetime in Multi-hop IoT networks. Even though clustering ensures an improved network lifetime, it cannot support for long time communications. Hence, this paper proposes a new and simple clustering mechanism based on Fuzzy C-Means (FCM) algorithm. Further, this work introduces a new energy provision mechanism for cluster heads through Hybrid Simultaneous Wireless Information and Power Transfer (H-SWIPT) mechanism. Unlike the conventional SWIPT which considers only a single source (past node’s RF signal), this H-SWIPT considers multiple sources namely past-hop node, sink node, co-channel interference, and neighbor nodes for energy harvesting. An extensive simulation experiment over the proposed mechanism proves its efficiency in terms of average energy consumption and network lifetime.

The NoCs are becoming an unavoidable solution of interconnection architecture in new systems that demand several performances like flexibility, extensibility, and less power consumption. The step of NoCs design is very important and there is a strong relation between the best conception and the optimal performances of NoCs. The mapping is one of the steps of NoCs design, classified as an NP-complete problem. In this paper, cat swarm optimization is exploited to the application mapping problem to minimize the communication cost. We compare the result obtained with some existing methods. The experimental results demonstrate the efficiency and effectiveness of cat swarm as we achieved a significant reduction in communication cost. For instance, for VOP and MPEG benchmarks, our algorithm saves more than 3.54% and 6.05% in communication cost compared to NMAP and previously proposed algorithms.

In recent years, there has been significant growth in applications of Unmanned Aerial Vehicles (UAVs). The demand for an autonomous UAV navigation is growing due to various applications in GPS-denied environments like disaster relief monitoring, search and rescue, mining, bridge inspections, space explorations, and military activities. Visual measurements possess a lot of accurate information which is extracted and exploited for UAV manoeuvring. This paper presents a comprehensive survey of vision-based UAV manoeuvring techniques. The approaches range from deep learning, digital elevation map, and optical flow to mathematical models. The outputs of these techniques cover the various aspects of autonomous navigation like velocity, thrust, yaw angle, heading angle, position, and height. The paper encompasses methods for both indoor and outdoor navigation. The techniques covered mange smooth UAV navigation in different and even unfavourable illumination conditions. Furthermore, this paper serves as a medium to gain insight into the essential aspects of drone navigation methods and their applications.

The Covid-19 pandemic has increased the global dependency on the internet. Millions of individuals use social networking sites to not only share information, but also their personal opinions. These facts and opinions are frequently unconfirmed, which result in the spread of incorrect information, generally alluded to as “Fake Content”. The most challenging aspect of social media is in determining the source of information. It's difficult to figure out who generated fake news once it's gone viral. Most available computational models have a key flaw in that they rely on the presence of inaccurate information to generate meaningful features, making disinformation mitigation measures difficult to predict. This paper presents a parallel approach to false information mitigation drawn from the field of Epidemiology using SIR(Susceptible, Infected, Recovered) to model the impact of fake data dissemination during Covid-19. SIR simulation is done using NetLogo in which the population is made up of two agents: Fake news believers and non-believers. To confirm our work, the concept of trust is also discussed which is a fundamental component of any fake news interaction. The level of trust can be expressed by assigning each node a pair of trust scores. We ran our experiments based on three common evaluation metrics: Accuracy, Precision, and Recall. The hybrid model shows an increase in accuracy by 81.4%, 77.1% , and 91.8% for the respective networks.

Analyzing data from the web is now one of the primary tasks, understood in a variety of manners and solved for a very wide variety of purposes. The talk describes the experience from a project, devoted to analyzing such data while drawing some more general conclusions. The project was aimed at distinguishing artificial ad-related traffic from the genuine one. The rationale is simple: The flow of money depends upon the number of clicks on/views of an ad. If so, fake clicking changes the market to the benefit of some, and to the loss of the other ones. The talk describes the problem and its conceptual framing, as well as a number of technical details, involving the issues and techniques of (1) variable analysis and choice; (2) clustering; (3) classification/classifiers; (4) potential hybrid techniques, along with citations of the most interesting results. These often imply definite general conclusions, some of them quite surprising.