Progress in Advanced Computing and Intelligent Engineering

Inhaltsverzeichnis

1. Frontmatter

2. Advanced Machine Learning Applications

   1. Frontmatter

   2. Prediction of Depression Using EEG: A Comparative Study

      Namrata P. Mohanty, Sweta Shree Dash, Sandeep Sobhan, Tripti Swarnkar

      Abstract

      The worldwide havoc of today’s world: depression, is increasing in this era. Depression is not any specific disease rather the determinant factor in the onset of numerous terrible diseases. With the increase in automation and artificial intelligence, it has become easier to predict depression before a much earlier time. The machine learning techniques are used in the classification of EEG for the prediction of different neuro-problems. EEG signals are the brain waves which can easily detect any abnormalities occurring in the brain waves, thereby making it easier to predict the seizure formation or depression. Proposed work uses the EEG signals for the analysis of brain waves, thereby predicting depression. In this paper, we have compared two widely used benchmark models, i.e., the k-NN and the ANN for the prediction of depression with an accuracy of 85%. This method will help doctors and medical associates in predicting diseases before the onset of its extreme phase, as well as assist them in providing the best treatments, possible in proper time.

   3. Prediction of Stroke Risk Factors for Better Pre-emptive Healthcare: A Public-Survey-Based Approach

      Debayan Banerjee, Jagannath Singh

      Abstract

      This work endeavours to explore the relation between certain behavioural traits and prevalent diseases among the sample population,reported in a public health survey,by means of machine learning techniques. Predictive models are developed to ascertain the significance statistically while also checking the fitness of the models to predict the diseases in a non-invasive way. Our study focuses on cardiovascular stroke from the BRFSS database of CDC, USA. The proposed model achieves 0.71 AuC in predicting stroke from purely behavioural features. Further analysis reveals an interesting behavioural trait: proper maintenance of an individual’s work–life balance, apart from the three main conventional habits: regular physical activity, healthy diet, abstinence from heavy smoking and drinking as the most significant actors for reducing the risk of potential stroke.

   4. Language Identification—A Supportive Tool for Multilingual ASR in Indian Perspective

      Basanta Kumar Swain, Sanghamitra Mohanty

      Abstract

      In this research paper, we have engineered a multilingual automatic speech recognition engine in Indian perspective by employing language identification technique. Our motherland, India is treated as land of many tongues. Our vision is leading to develop a system that would aid in man–machine communication in Indian spoken languages. We have developed two vital models, namely language identification and speech recognition using array of pattern recognition techniques, viz. k-NN, SVM and HMM. We have tackled both LID and multilingual ASR tasks over three Indian spoken languages, namely Odia, Hindi and Indian English. We have experimented over short as well as long durational isolated spoken words for LID task purpose. Finally, we have integrated LID module with multilingual ASR in order to transcribe the words in desired spoken language.

   5. Ensemble Methods to Predict the Locality Scope of Indian and Hungarian Students for the Real Time: Preliminary Results

      Chaman Verma, Zoltán Illés, Veronika Stoffová

      Abstract

      In the present study, we presented ensemble classifier to predict the locality scope (National or International) of the student based on their motherland and sex toward Information and Communication Technology (ICT) and Mobile Technology (MT). For this, a primary dataset of 331 samples from Indian and Hungarian university was gathered during the academic year 2017–2018. The dataset contained 331 instances and 37 features which belonged to the four major ICT parameters attitude, development and availability, educational benefits and usability of modern ICT resources, and mobile technology in higher education. In addition to class balancing with Synthetic Minority Over-Sampling Technique (SMOTE), Adaptive Boosting (AdboostM1) and bagging ensemble technique is applied with Artificial Neural Network (ANN) and Random Forest (RF) classifiers in Weka tool. Findings of the study infer that the ANN achieved higher accuracy (92.94%) as compared to RF’s accuracy (92.25%). The author’s contribution is to apply ensemble methods with standard classifiers to provide more accurate and consistent results. On the one hand, with the use of bagging, the ANN achieved 92.94% accuracy, and on the other hand, AdboostM1 has also significantly improved the prediction accuracy and RF provided 92.25% accuracy. Further, the statistical T-test at the 0.05 significance level proved no significant difference between the accuracy of RF and ANN classifier to predict the locality scope of the student. Also, the authors found a significant difference between the CPU prediction time between bagging with ANN and AdboostM1 with RF.

   6. Automatic Detection and Classification of Tomato Pests Using Support Vector Machine Based on HOG and LBP Feature Extraction Technique

      Gayatri Pattnaik, K. Parvathi

      Abstract

      The automatic detection and classification of insect pest is emerged as one of the interesting research areas in agriculture sector to ensure reduction of damages due to pest. From the general process of detection of pest, feature extraction plays a significant role. It extracts features from the segmented image obtained by segmentation process, and then extracted images are being transferred to a classifier for the operations. In this work, we studied and implemented two feature extraction techniques, i.e., Histogram of Oriented Gradient (HOG) and Local Binary Pattern techniques (LBP). The comparison result expressed that HOG performs better than its counterpart. The result comes with accuracy of 97% for HOG. Here, we are adopting SVM-based pest classification as a test case.

   7. Poly Scale Space Technique for Feature Extraction in Lip Reading: A New Strategy

      M. S. Nandini, Nagappa U. Bhajantri, Trisiladevi C. Nagavi

      Abstract

      Lip reading involves the extraction of visual speech information contained in the inner and outer lip contour. Visibility of teeth and tongue during speech provides important speech cues. Particularly for fricatives, the place of articulation can often be determined visually, that is, for labiodentals (upper teeth or lower lip), interdentals (behind tongue or front teeth), and alveolar (tongue touching gun ridge). Other speech information might be contained in the protrusion and wrinkling of lips. Feature extraction is a remarkable process in lip reading as it holds an important role in lip reading classification. In Improved Speeded Up Robust Feature (ISURF), extraction for finding an exact edge is difficult because of more false corner ratio. In case of PSST, exact edges could be obtained with reduced false corner ratio. This paper provides the idea about PSST based on Harris algorithm and gives more precise edge detection in different illumination conditions.

   8. Machine Learning Methods for Vehicle Positioning in Vehicular Ad-Hoc Networks

      Suryakanta Nayak, Partha Sarathi Das, Satyasen Panda

      Abstract

      Unambiguous vehicular sensing is one of the most important aspects in autonomous driving in vehicular ad-hoc networks. The conventional techniques such as communication-based technologies (e.g., GPS) or the reflection-based technologies (e.g., RADAR, LIDAR) have various limitations in detecting concealed vehicles in dense urban areas without line of sight which may trigger serious accidents for autonomous vehicles. To address this issue, this paper proposed a machine learning method based on stochastic Gaussian process regression (SGP) to position vehicles in a distributed vehicular system with received signal vector (RSV) information. To estimate the test vehicle position and respective position errors, the proposed SGP method records the RSV readings at neighboring locations with continuous approximation of the vehicle-to-vehicle (V2V) distance, angle of arrival (AoA), and path delay. Then, the subsequent averaging of the training RSVs minimizes the effects of the shadowing noise and multipath fading. The prediction performance of the proposed learning approach is measured in terms of the root mean square prediction error (RMSE) in a realistic environment. Finally, the prediction performance of the proposed learning method is compared with other existing fingerprinting methods for error-free location estimation of the vehicular network.

   9. Effectiveness of Swarm-Based Metaheuristic Algorithm in Data Classification Using Pi-Sigma Higher Order Neural Network

      Nibedan Panda, Santosh Kumar Majhi

      Abstract

      In this paper, Salp Swarm Algorithm (SSA) is employed in training the Higher Order Neural Network (HONN) for data classification task. In machine learning approach, to train artificial neural network is considered a difficult task which gains the attention of researchers recently. The difficulty of Artificial Neural Networks (ANNs) arises due to its nonlinearity nature and unknown set of initial parameters. Traditional training algorithms exhibit poor performance in terms of local optima avoidance and convergence rate, for which metaheuristic based optimization emerges as a suitable alternative. The performance of the proposed SSA-based HONN method has been verified by considering various classification measures over benchmark datasets chosen from UCI repository and the outcome obtained by the said method is compared with the state-of-art evolutionary algorithms. From the outcome reported, the proposed method outperforms over the recent algorithms which confirm its supremacy in terms of better exploration and exploitation capability.

   10. Deep Learning for Cover Song Apperception

       D. Khasim Vali, Nagappa U. Bhajantri

       Abstract

       In this work, we proposed a cover song recognition system using deep learning. From the literature, understand that most of the works extract the discriminate feature that classifies the cover song between a pair of songs and calculates the dissimilarity or similarity between the two songs based on the observation, which is a meaningful pattern between cover songs. Moreover, it inspires reformulating the cover song apperception obstacle in a machine learning framework. In other words, essentially builds the cover song recognition system using Convolution Neural Network (CNN) and Mel Frequency Cepstral Coefficients (MFCCs) features following the construction of the data set composed of cover song pairs. The prepared CNN yields the likelihood of being in the spread tune connection given a cross-closeness grid produced from any two bits of music and recognizes the spread tune by positioning on the likelihood. Test results display the prescribed methodology that has accomplished enhanced execution tantamount to the cutting edge endeavors.

   11. SVM-Based Drivers Drowsiness Detection Using Machine Learning and Image Processing Techniques

       P. Rasna, M. B. Smithamol

       Abstract

       In this paper, we propose an efficient algorithm for driver drowsiness detection and efficient alert system. The existing works mainly follow vehicle-based measures, physiological-based measures, behavioral-based measures. Moreover, the works based on behavioral measures mainly focused on eye movements, yawning, and head position. The proposed method uses more relevant and appropriate behavioral features such as significant variation in aspect ratio of eyes, mouth opening ratio, nose length bending, and the changes that happened in eyebrows, wrinkles, ear due to drowsiness. The binary SVM classifier is used for classification whether the driver is drowsy or not. The inclusion of these features helped in developing more efficient driver drowsiness detection system. The proposed system shows 97.5% accuracy and 97.8% detection rate.

   12. Fusion of Artificial Intelligence for Multidisciplinary Optimization: Skidding Track—Case Study

       Abhishek Nigam, Debi Prasad Ghosh

       Abstract

       The demand for multidisciplinary optimized products in industrial engineering is high due to growing competition. But it is a highly expensive and time-consuming process. The integration of artificial intelligence in engineering design is an area of high demand. So here, using the fusion of AI, a method has been proposed to solve a multidisciplinary problem using the case study of skidding tracks.

   13. A Single Document Assamese Text Summarization Using a Combination of Statistical Features and Assamese WordNet

       Nomi Baruah, Shikhar Kr. Sarma, Surajit Borkotokey

       Abstract

       In this paper, an extractive text summarization approach using Assamese WordNet is proposed, and the difficulties faced while extracting summary in the Assamese document are discussed. The Assamese language is a low-level language. Synset is applied from Assamese WordNet. The various features used for identifying the most salient sentences to generate effective summary aspects such as TF-IDF, sentence length, sentence position and numerical identification are considered. Automatic Text Summarization in the Assamese language is still in an early stage and this language does not have its own approach. So, the text summarization approach is compared to the approaches applied in Bengali and Bangla language approaches as these languages share a script that is quite similar having slight variations in certain letters. The effectiveness of our proposed approach is demonstrated through a set of experiments carried out using ROUGE measure, and the evaluation is depicted in terms of Precision, Recall and F1-score.

   14. SVM and Ensemble-SVM in EEG-Based Person Identification

       Banee Bandana Das, Saswat Kumar Ram, Bibudhendu Pati, Chhabi Rani Panigrahi, Korra Sathya Babu, Ramesh Kumar Mohapatra

       Abstract

       Biometric person identification is getting more effective and popular because of Electroencephalography (EEG). EEG signals can be captured from human scalp invasively or non-invasively with the help of electrodes. EEG-based biometric system is more secure and unique for person identification. In this paper, we have used two different states to explore the adaptive and uniqueness of the EEG-based biometric system. We have used Eyes Open (EO) state as well as Eyes Closed (EC) state of a EEG motor imagery publicly available dataset of 109 users.The model is trained and tested with EO and EC states alternatively to prove the reliability and robustness of the model. The biometric person identification model has been designed using Support Vector machine (SVM) for classification. We achieved a notable person identification rate of 96% (EO) and 91.78% (EC) using SVM with Radial Basis Function (RBF) kernel. We have also used Ensemble Support Vector Machine (ESVM) to enhance the performance of person identification and observed the average performance accuracy of 96.16% with n number of classifier.

   15. A Self-Acting Mechanism to Engender Highlights of a Tennis Game

       Ramanathan Arunachalam, Abishek Kumar

       Abstract

       With the rapid growth of digital media content specifically in Videos, different genre of sports matches are among the foremost event of the Television industry. Currently, any live recording goes into a RAID system of sorts, where all the footage from various cameras is recorded and kept available for random access. The proposed work aims to achieve automation in the broadcasting industry by creating a Self-Acting application which generates the highlights package without any external people devoted to this cause.

   16. Performance Evaluation of RF and SVM for Sugarcane Classification Using Sentinel-2 NDVI Time-Series

       Shyamal Virnodkar, V. K. Pachghare, V. C. Patil, Sunil Kumar Jha

       Abstract

       Sentinel-2 optical time-series images obtained at high resolution are creditable for cropland mapping which is the key for sustainable agriculture. The presented work was conducted in a heterogeneous region in Sameerwadi with an aim to classify sugarcane crops, with mainly two groups so as to provide a sugarcane field map, using Sentinel-2 normalized difference vegetation index (NDVI) time-series data. The potential of two better-known machine learning (ML) classifiers, random forest (RF) and support vector machine (SVM), was investigated to identify seven classes including sugarcane, early sugarcane, maize, waterbody, fallow land, built-up and bare land, and a sugarcane crop map is produced. Both the classifiers were able to effectively classify sugarcane areas and other land covers from the time-series data. Our results show that RF achieved higher overall accuracy (88.61%) than SVM having an overall accuracy of 81.86%. This study demonstrated that utilizing the Sentinel-2 NDVI time-series with RF and SVM successfully classified sugarcane crop fields.

   17. Classification of Nucleotides Using Memetic Algorithms and Computational Methods

       Rajesh Eswarawaka, S. Venkata Suryanarayana, Purnachand Kollapudi, Mrutyunjaya S. Yalawar

       Abstract

       This paper presents an approach to solve an optimization problem using clustering by genetic algorithm approach. The central idea is to form clusters of patients’ nucleotide data sets. The genetic algorithm is applied to this initial cluster population. The fitness function for the genetic algorithm is calculated using intra-cluster and inter-cluster distances. Later genetic crossover functions are applied. This procedure is iterated until the stopping condition is reached. The superiority of this algorithm lies in comparing the performance with Ant Colony Optimization and simulated annealing algorithms.

   18. A Novel Approach to Detect Emergency Using Machine Learning

       Sarmistha Nanda, Chhabi Rani Panigrahi, Bibudhendu Pati, Abhishek Mishra

       Abstract

       Human activity is always a reflection of its external environmental conditions. If a group of people is in some emergency, then their activities and behaviour will be different as compared to normal conditions. To detect an emergency, Human Activity Recognition (HAR) can play an important role. Human activities such as shouting, running here and there, crying, searching for an exit door can be taken into consideration as an emergency indicator. By detecting the emergency and its degree, the Emergency Management System (EMS) can manage the situation efficiently. In this work, we use machine learning algorithms such as Random Forest (RF), IBK, Bagging, J48 and MLP on WISDM Smartphone and Smartwatch Activity and Biometric Dataset for human activity recognition and RF is found to be the best algorithm with classification accuracy 87.1977% among all other considered techniques.