Search results for: Imbalanced dataset

Authors: E. Tchandao Mangamana, V. Cariou, E. Vigneau, R. Glele Kakai, E. M. Qannari

Abstract:

A new definition of a latent variable associated with a dataset makes it possible to propose variants of the PLS2 regression and the multi-block PLS (MB-PLS). We shall refer to these variants as Rd-PLS regression and Rd-MB-PLS respectively because they are inspired by both Redundancy analysis and PLS regression. Usually, a latent variable t associated with a dataset Z is defined as a linear combination of the variables of Z with the constraint that the length of the loading weights vector equals 1. Formally, t=Zw with ‖w‖=1. Denoting by Z' the transpose of Z, we define herein, a latent variable by t=ZZ’q with the constraint that the auxiliary variable q has a norm equal to 1. This new definition of a latent variable entails that, as previously, t is a linear combination of the variables in Z and, in addition, the loading vector w=Z’q is constrained to be a linear combination of the rows of Z. More importantly, t could be interpreted as a kind of projection of the auxiliary variable q onto the space generated by the variables in Z, since it is collinear to the first PLS1 component of q onto Z. Consider the situation in which we aim to predict a dataset Y from another dataset X. These two datasets relate to the same individuals and are assumed to be centered. Let us consider a latent variable u=YY’q to which we associate the variable t= XX’YY’q. Rd-PLS consists in seeking q (and therefore u and t) so that the covariance between t and u is maximum. The solution to this problem is straightforward and consists in setting q to the eigenvector of YY’XX’YY’ associated with the largest eigenvalue. For the determination of higher order components, we deflate X and Y with respect to the latent variable t. Extending Rd-PLS to the context of multi-block data is relatively easy. Starting from a latent variable u=YY’q, we consider its ‘projection’ on the space generated by the variables of each block Xk (k=1, ..., K) namely, tk= XkXk'YY’q. Thereafter, Rd-MB-PLS seeks q in order to maximize the average of the covariances of u with tk (k=1, ..., K). The solution to this problem is given by q, eigenvector of YY’XX’YY’, where X is the dataset obtained by horizontally merging datasets Xk (k=1, ..., K). For the determination of latent variables of order higher than 1, we use a deflation of Y and Xk with respect to the variable t= XX’YY’q. In the same vein, extending Rd-MB-PLS to the path modeling setting is straightforward. Methods are illustrated on the basis of case studies and performance of Rd-PLS and Rd-MB-PLS in terms of prediction is compared to that of PLS2 and MB-PLS.

Keywords: multiblock data analysis, partial least squares regression, path modeling, redundancy analysis

Procedia PDF Downloads 110

Authors: Shraddha Chaudhary, Raksha Agarwal, Niladri Chatterjee

Abstract:

This work demonstrates a web crawler-based generalized end-to-end open domain Question Answering (QA) system. An efficient QA system requires a significant amount of domain knowledge to answer any question with the aim to find an exact and correct answer in the form of a number, a noun, a short phrase, or a brief piece of text for the user's questions. Analysis of the question, searching the relevant document, and choosing an answer are three important steps in a QA system. This work uses a web scraper (Beautiful Soup) to extract K-documents from the web. The value of K can be calibrated on the basis of a trade-off between time and accuracy. This is followed by a passage ranking process using the MS-Marco dataset trained on 500K queries to extract the most relevant text passage, to shorten the lengthy documents. Further, a QA system is used to extract the answers from the shortened documents based on the query and return the top 3 answers. For evaluation of such systems, accuracy is judged by the exact match between predicted answers and gold answers. But automatic evaluation methods fail due to the linguistic ambiguities inherent in the questions. Moreover, reference answers are often not exhaustive or are out of date. Hence correct answers predicted by the system are often judged incorrect according to the automated metrics. One such scenario arises from the original Google Natural Question (GNQ) dataset which was collected and made available in the year 2016. Use of any such dataset proves to be inefficient with respect to any questions that have time-varying answers. For illustration, if the query is where will be the next Olympics? Gold Answer for the above query as given in the GNQ dataset is “Tokyo”. Since the dataset was collected in the year 2016, and the next Olympics after 2016 were in 2020 that was in Tokyo which is absolutely correct. But if the same question is asked in 2022 then the answer is “Paris, 2024”. Consequently, any evaluation based on the GNQ dataset will be incorrect. Such erroneous predictions are usually given to human evaluators for further validation which is quite expensive and time-consuming. To address this erroneous evaluation, the present work proposes an automated approach for evaluating time-dependent question-answer pairs. In particular, it proposes a metric using the current timestamp along with top-n predicted answers from a given QA system. To test the proposed approach GNQ dataset has been used and the system achieved an accuracy of 78% for a test dataset comprising 100 QA pairs. This test data was automatically extracted using an analysis-based approach from 10K QA pairs of the GNQ dataset. The results obtained are encouraging. The proposed technique appears to have the possibility of developing into a useful scheme for gathering precise, reliable, and specific information in a real-time and efficient manner. Our subsequent experiments will be guided towards establishing the efficacy of the above system for a larger set of time-dependent QA pairs.

Keywords: web-based information retrieval, open domain question answering system, time-varying QA, QA evaluation

Procedia PDF Downloads 74

Authors: Shakila N. Senarath, Dinesh Asanka, Janaka Wijayanayake

Abstract:

The necessity of cosmetic products is arising to fulfill consumer needs of personality appearance and hygiene. A cosmetic product consists of various chemical ingredients which may help to keep the skin healthy or may lead to damages. Every chemical ingredient in a cosmetic product does not perform on every human. The most appropriate way to select a healthy cosmetic product is to identify the texture of the body first and select the most suitable product with safe ingredients. Therefore, the selection process of cosmetic products is complicated. Consumer surveys have shown most of the time, the selection process of cosmetic products is done in an improper way by consumers. From this study, a content-based system is suggested that recommends cosmetic products for the human factors. To such an extent, the skin type, gender and price range will be considered as human factors. The proposed system will be implemented by using Machine Learning. Consumer skin type, gender and price range will be taken as inputs to the system. The skin type of consumer will be derived by using the Baumann Skin Type Questionnaire, which is a value-based approach that includes several numbers of questions to derive the user’s skin type to one of the 16 skin types according to the Bauman Skin Type indicator (BSTI). Two datasets are collected for further research proceedings. The user data set was collected using a questionnaire given to the public. Those are the user dataset and the cosmetic dataset. Product details are included in the cosmetic dataset, which belongs to 5 different kinds of product categories (Moisturizer, Cleanser, Sun protector, Face Mask, Eye Cream). An alternate approach of TF-IDF (Term Frequency – Inverse Document Frequency) is applied to vectorize cosmetic ingredients in the generic cosmetic products dataset and user-preferred dataset. Using the IF-IPF vectors, each user-preferred products dataset and generic cosmetic products dataset can be represented as sparse vectors. The similarity between each user-preferred product and generic cosmetic product will be calculated using the cosine similarity method. For the recommendation process, a similarity matrix can be used. Higher the similarity, higher the match for consumer. Sorting a user column from similarity matrix in a descending order, the recommended products can be retrieved in ascending order. Even though results return a list of similar products, and since the user information has been gathered, such as gender and the price ranges for product purchasing, further optimization can be done by considering and giving weights for those parameters once after a set of recommended products for a user has been retrieved.

Keywords: content-based filtering, cosmetics, machine learning, recommendation system

Procedia PDF Downloads 108

Authors: Esha Banerjee, Atul Kumar Ojha, Girish Nath Jha

Abstract:

This study aims to develop an intonation labeled database for Hindi. Although no single standard for prosody labeling exists in Hindi, researchers in the past have employed perceptual and statistical methods in literature to draw inferences about the behavior of prosody patterns in Hindi. Based on such existing research and largely agreed upon intonational theories in Hindi, this study attempts to develop a manually annotated prosodic corpus of Hindi speech data, which can be used for training speech models for natural-sounding speech in the future. 100 sentences ( 500 words) each for declarative and interrogative types have been labeled using Praat.

Keywords: speech dataset, Hindi, intonation, labeled corpus

Procedia PDF Downloads 159

Authors: Gehad S. Kaseb, Mona F. Ahmed

Abstract:

Arabic Sentiment Analysis (SA) is one of the most common research fields with many open areas. Few studies apply SA to Arabic dialects. This paper proposes different pre-processing steps and a modified methodology to improve the accuracy using normal Support Vector Machine (SVM) classification. The paper works on two datasets, Arabic Sentiment Tweets Dataset (ASTD) and Extended Arabic Tweets Sentiment Dataset (Extended-AATSD), which are publicly available for academic use. The results show that the classification accuracy approaches 86%.

Keywords: Arabic, classification, sentiment analysis, tweets

Procedia PDF Downloads 113

Authors: Yash Jain

Abstract:

The US Center for Disease Control has recommended wearing masks to slow the spread of the virus. The research uses a video feed from a camera to conduct real-time classifications of whether or not a human is correctly wearing a mask, incorrectly wearing a mask, or not wearing a mask at all. Utilizing two distinct datasets from the open-source website Kaggle, a mask detection network had been trained. The first dataset that was used to train the model was titled 'Face Mask Detection' on Kaggle, where the dataset was retrieved from and the second dataset was titled 'Face Mask Dataset, which provided the data in a (YOLO Format)' so that the TinyYoloV3 model could be trained. Based on the data from Kaggle, two machine learning models were implemented and trained: a Tiny YoloV3 Real-time model and a two-stage neural network classifier. The two-stage neural network classifier had a first step of identifying distinct faces within the image, and the second step was a classifier to detect the state of the mask on the face and whether it was worn correctly, incorrectly, or no mask at all. The TinyYoloV3 was used for the live feed as well as for a comparison standpoint against the previous two-stage classifier and was trained using the darknet neural network framework. The two-stage classifier attained a mean average precision (MAP) of 80%, while the model trained using TinyYoloV3 real-time detection had a mean average precision (MAP) of 59%. Overall, both models were able to correctly classify stages/scenarios of no mask, mask, and incorrectly worn masks.

Keywords: datasets, classifier, mask-detection, real-time, TinyYoloV3, two-stage neural network classifier

Procedia PDF Downloads 127

Authors: Teklu Urgessa, Wookjae Maeng, Joong Seek Lee

Abstract:

Application of five implementations of three data mining classification techniques was experimented for extracting important insights from tourism data. The aim was to find out the best performing algorithm among the compared ones for tourism knowledge discovery. Knowledge discovery process from data was used as a process model. 10-fold cross validation method is used for testing purpose. Various data preprocessing activities were performed to get the final dataset for model building. Classification models of the selected algorithms were built with different scenarios on the preprocessed dataset. The outperformed algorithm tourism dataset was Random Forest (76%) before applying information gain based attribute selection and J48 (C4.5) (75%) after selection of top relevant attributes to the class (target) attribute. In terms of time for model building, attribute selection improves the efficiency of all algorithms. Artificial Neural Network (multilayer perceptron) showed the highest improvement (90%). The rules extracted from the decision tree model are presented, which showed intricate, non-trivial knowledge/insight that would otherwise not be discovered by simple statistical analysis with mediocre accuracy of the machine using classification algorithms.

Keywords: classification algorithms, data mining, knowledge discovery, tourism

Procedia PDF Downloads 267

Authors: Megha Gupta, Nupur Prakash

Abstract:

Identification of plant diseases has been performed using machine learning and deep learning models on the datasets containing images of healthy and diseased plant leaves. The current study carries out an evaluation of some of the deep learning models based on convolutional neural network (CNN) architectures for identification of plant diseases. For this purpose, the publicly available New Plant Diseases Dataset, an augmented version of PlantVillage dataset, available on Kaggle platform, containing 87,900 images has been used. The dataset contained images of 26 diseases of 14 different plants and images of 12 healthy plants. The CNN models selected for the study presented in this paper are AlexNet, ZFNet, VGGNet (four models), GoogLeNet, and ResNet (three models). The selected models are trained using PyTorch, an open-source machine learning library, on Google Colaboratory. A comparative study has been carried out to analyze the high degree of accuracy achieved using these models. The highest test accuracy and F1-score of 99.59% and 0.996, respectively, were achieved by using GoogLeNet with Mini-batch momentum based gradient descent learning algorithm.

Keywords: comparative analysis, convolutional neural networks, deep learning, plant disease identification

Procedia PDF Downloads 163

Authors: Seyyedeh Faezeh Hassani Ziabari, Sadegh Eskandari, Maziar Salahi

Abstract:

Infinite Feature Selection (IFS) algorithm is an efficient feature selection algorithm that selects a subset of features of all sizes (including infinity). In this paper, we present an improved version of it, called clustering IFS (CIFS), by clustering the dataset in advance. To do so, first, we apply the K-means algorithm to cluster the dataset, then we apply IFS. In the CIFS method, the spatial and temporal complexities are reduced compared to the IFS method. Experimental results on 6 datasets show the superiority of CIFS compared to IFS in terms of accuracy, running time, and memory consumption.

Keywords: feature selection, infinite feature selection, clustering, graph

Procedia PDF Downloads 94

Authors: Kaleeswari R. K., Seevagan L .

Abstract:

Soil quality in an agroecosystem is influenced by the cropping system, water and soil fertility management. A valid soil quality index would help to assess the soil and crop management practices for desired productivity and soil health. The soil quality indices also provide an early indication of soil degradation and needy remedial and rehabilitation measures. Imbalanced fertilization and inadequate organic carbon dynamics deteriorate soil quality in an intensive cropping system. The rice soil ecosystem is different from other arable systems since rice is grown under submergence, which requires a different set of key soil attributes for enhancing soil quality and productivity. Assessment of the soil quality index involves indicator selection, indicator scoring and comprehensive score into one index. The most appropriate indicator to evaluate soil quality can be selected by establishing the minimum data set, which can be screened by linear and multiple regression factor analysis and score function. This investigation was carried out in intensive rice cultivating regions (having >1.0 lakh hectares) of Tamil Nadu viz., Thanjavur, Thiruvarur, Nagapattinam, Villupuram, Thiruvannamalai, Cuddalore and Ramanathapuram districts. In each district, intensive rice growing block was identified. In each block, two sampling grids (10 x 10 sq.km) were used with a sampling depth of 10 – 15 cm. Using GIS coordinates, and soil sampling was carried out at various locations in the study area. The number of soil sampling points were 41, 28, 28, 32, 37, 29 and 29 in Thanjavur, Thiruvarur, Nagapattinam, Cuddalore, Villupuram, Thiruvannamalai and Ramanathapuram districts, respectively. Principal Component Analysis is a data reduction tool to select some of the potential indicators. Principal Component is a linear combination of different variables that represents the maximum variance of the dataset. Principal Component that has eigenvalues equal or higher than 1.0 was taken as the minimum data set. Principal Component Analysis was used to select the representative soil quality indicators in rice soils based on factor loading values and contribution percent values. Variables having significant differences within the production system were used for the preparation of the minimum data set. Each Principal Component explained a certain amount of variation (%) in the total dataset. This percentage provided the weight for variables. The final Principal Component Analysis based soil quality equation is SQI = ∑ i=1 (W ᵢ x S ᵢ); where S- score for the subscripted variable; W-weighing factor derived from PCA. Higher index scores meant better soil quality. Soil respiration, Soil available Nitrogen and Potentially Mineralizable Nitrogen were assessed as soil quality indicators in rice soil of the Cauvery Delta zone covering Thanjavur, Thiruvavur and Nagapattinam districts. Soil available phosphorus could be used as a soil quality indicator of rice soils in the Cuddalore district. In rain-fed rice ecosystems of coastal sandy soil, DTPA – Zn could be used as an effective soil quality indicator. Among the soil parameters selected from Principal Component Analysis, Microbial Biomass Nitrogen could be used quality indicator for rice soils of the Villupuram district. Cauvery Delta zone has better SQI as compared with other intensive rice growing zone of Tamil Nadu.

Keywords: soil quality index, soil attributes, soil mapping, and rice soil

Procedia PDF Downloads 53

Authors: M. Novak, E. Munoz, A. Jana, M. Nelemans

Abstract:

Cities and local governments the world over are interested to employ circular strategies as a means to bring about food security, create employment and increase resilience. The selection and implementation of circular strategies is facilitated by modeling the effects of strategies locally and understanding the impacts such strategies have had in other (comparable) cities and how that would translate locally. Urban areas are heterogeneous because of their geographic, economic, social characteristics, governance, and culture. In order to better understand the effect of circular strategies on urban systems, we create a dataset for over 300 cities around the world designed to facilitate circular strategy scenario modeling. This new dataset integrates data from over 20 prominent global national and urban data sources, such as the Global Human Settlements layer and International Labour Organisation, as well as incorporating employment data from over 150 cities collected bottom up from local departments and data providers. The dataset is made to be reproducible. Various clustering techniques are explored in the paper. The result is sets of clusters of cities, which can be used for further research, analysis, and support comparative, regional, and national policy making on circular cities.

Keywords: data integration, urban innovation, cluster analysis, circular economy, city profiles, scenario modelling

Procedia PDF Downloads 154

Authors: Brock Buckingham, Zhe Lin, Wenxuan Guo

Abstract:

Accurate elevation data is critical in deriving topographic attributes for the precision management of crop inputs, especially water and nutrients. Traditional ground-based elevation data acquisition is time consuming, labor intensive, and often inconvenient at the field scale. Various unmanned aerial systems (UAS) provide the capability of generating digital elevation data from high-resolution images. The objective of this study was to compare the performance of two UAS with different global positioning system (GPS) receivers in determining elevation at the field scale. A DJI Phantom 4 Pro and a DJI Phantom 4 RTK(real-time kinematic) were applied to acquire images at three heights, including 40m, 80m, and 120m above ground. Forty ground control panels were placed in the field, and their geographic coordinates were determined using an RTK GPS survey unit. For each image acquisition using a UAS at a particular height, two elevation datasets were generated using the Pix4D stitching software: a calibrated dataset using the surveyed coordinates of the ground control panels and an uncalibrated dataset without using the surveyed coordinates of the ground control panels. Elevation values for each panel derived from the elevation model of each dataset were compared to the corresponding coordinates of the ground control panels. The coefficient of the determination (R²) and the root mean squared error (RMSE) were used as evaluation metrics to assess the performance of each image acquisition scenario. RMSE values for the uncalibrated elevation dataset were 26.613 m, 31.141 m, and 25.135 m for images acquired at 120 m, 80 m, and 40 m, respectively, using the Phantom 4 Pro UAS. With calibration for the same UAS, the accuracies were significantly improved with RMSE values of 0.161 m, 0.165, and 0.030 m, respectively. The best results showed an RMSE of 0.032 m and an R² of 0.998 for calibrated dataset generated using the Phantom 4 RTK UAS at 40m height. The accuracy of elevation determination decreased as the flight height increased for both UAS, with RMSE values greater than 0.160 m for the datasets acquired at 80 m and 160 m. The results of this study show that calibration with ground control panels improves the accuracy of elevation determination, especially for the UAS with a regular GPS receiver. The Phantom 4 Pro provides accurate elevation data with substantial surveyed ground control panels for the 40 m dataset. The Phantom 4 Pro RTK UAS provides accurate elevation at 40 m without calibration for practical precision agriculture applications. This study provides valuable information on selecting appropriate UAS and flight heights in determining elevation for precision agriculture applications.

Keywords: unmanned aerial system, elevation, precision agriculture, real-time kinematic (RTK)

Procedia PDF Downloads 133

Authors: Ulrich Wake, Eniman Syamsuddin

Abstract:

The model for camera model identificaiton is trained using pretrained model ResNet43 and ResNet50. The dataset consists of 500 photos of each phone. Dataset is divided into 1280 photos for training, 320 photos for validation and 400 photos for testing. The model is trained using One Cycle Policy Method and tested using Test-Time Augmentation. Furthermore, the model is trained for 50 epoch using regularization such as drop out and early stopping. The result is 90% accuracy for validation set and above 85% for Test-Time Augmentation using ResNet50. Every model is also trained by slightly updating the pretrained model’s weights

Keywords: ​ One Cycle Policy, ResNet34, ResNet50, Test-Time Agumentation

Procedia PDF Downloads 169

Authors: Insaf Ajili, Malik Mallem, Jean-Yves Didier

Abstract:

Human motion recognition has been extensively increased in recent years due to its importance in a wide range of applications, such as human-computer interaction, intelligent surveillance, augmented reality, content-based video compression and retrieval, etc. However, it is still regarded as a challenging task especially in realistic scenarios. It can be seen as a general machine learning problem which requires an effective human motion representation and an efficient learning method. In this work, we introduce a descriptor based on Laban Movement Analysis technique, a formal and universal language for human movement, to capture both quantitative and qualitative aspects of movement. We use Discrete Hidden Markov Model (DHMM) for training and classification motions. We improve the classification algorithm by proposing two DHMMs for each motion class to process the motion sequence in two different directions, forward and backward. Such modification allows avoiding the misclassification that can happen when recognizing similar motions. Two experiments are conducted. In the first one, we evaluate our method on a public dataset, the Microsoft Research Cambridge-12 Kinect gesture data set (MSRC-12) which is a widely used dataset for evaluating action/gesture recognition methods. In the second experiment, we build a dataset composed of 10 gestures(Introduce yourself, waving, Dance, move, turn left, turn right, stop, sit down, increase velocity, decrease velocity) performed by 20 persons. The evaluation of the system includes testing the efficiency of our descriptor vector based on LMA with basic DHMM method and comparing the recognition results of the modified DHMM with the original one. Experiment results demonstrate that our method outperforms most of existing methods that used the MSRC-12 dataset, and a near perfect classification rate in our dataset.

Keywords: human motion recognition, motion representation, Laban Movement Analysis, Discrete Hidden Markov Model

Procedia PDF Downloads 175

Authors: Shamm Kidane, Ibrahim Abdella, Fitsum Gaim, Simon Mulugeta, Sirak Asmerom, Natnael Ambasager, Yoel Ghebrihiwot

Abstract:

The call for natural language processing (NLP) systems for low-resource languages has become more apparent than ever in the past few years, with the arduous challenges still present in preparing such systems. This paper presents an improved dataset version of the Nagaoka Tigrinya Corpus for Parts-of-Speech (POS) classification system in the Tigrinya language. The size of the initial Nagaoka dataset was incremented, totaling the new tagged corpus to 118K tokens, which comprised the 12 basic POS annotations used previously. The additional content was also annotated manually in a stringent manner, followed similar rules to the former dataset and was formatted in CONLL format. The system made use of the novel approach in NLP tasks and use of the monolingually pre-trained TiELECTRA, TiBERT and TiRoBERTa transformer models. The highest achieved score is an impressive weighted F1-score of 94.2%, which surpassed the previous systems by a significant measure. The system will prove useful in the progress of NLP-related tasks for Tigrinya and similarly related low-resource languages with room for cross-referencing higher-resource languages.

Keywords: Tigrinya POS corpus, TiBERT, TiRoBERTa, conditional random fields

Procedia PDF Downloads 59

Authors: Saba Ebrahimi, Hedieh Ashrafi

Abstract:

Predicting the cover type of forests is a challenge for natural resource managers. In this project, we aim to perform a comprehensive comparative study of two well-known classification methods, support vector machine (SVM) and decision tree (DT). The comparison is first performed among different types of each classifier, and then the best of each classifier will be compared by considering different evaluation metrics. The effect of boosting and bagging for decision trees is also explored. Furthermore, the effect of principal component analysis (PCA) and feature selection is also investigated. During the project, the forest cover-type dataset from the remote sensing and GIS program is used in all computations.

Keywords: classification methods, support vector machine, decision tree, forest cover-type dataset

Procedia PDF Downloads 178

Authors: Shimaa Holail, Tamer Saleh, Xiongwu Xiao

Abstract:

Building change detection (BCD) from satellite imagery is an essential topic in urbanization monitoring, agricultural land management, and updating geospatial databases. Recently, methods for detecting changes based on deep learning have made significant progress and impressive results. However, it has the problem of being insensitive to changes in buildings with complex spectral differences, and the features being extracted are not discriminatory enough, resulting in incomplete buildings and irregular boundaries. To overcome these problems, we propose a dual Siamese network based on the Unet model with the addition of a deep supervision strategy (DS) in this paper. This network consists of a backbone (encoder) based on ImageNet pre-training, a fusion block, and feature pyramid networks (FPN) to enhance the step-by-step information of the changing regions and obtain a more accurate BCD map. To train the proposed method, we created a new dataset (EGY-BCD) of high-resolution and multi-temporal aerial images captured over New Cairo in Egypt to detect building changes for this purpose. The experimental results showed that the proposed method is effective and performs well with the EGY-BCD dataset regarding the overall accuracy, F1-score, and mIoU, which were 91.6 %, 80.1 %, and 73.5 %, respectively.

Keywords: building change detection, deep supervision, semantic segmentation, EGY-BCD dataset

Procedia PDF Downloads 74

Authors: Harshan Baskar, Anirudh S. Chakravarthy, Prateek Garg, Divyam Goel, Abhijith S. Raj, Kshitij Kumar, Lakshya, Ravichandra Parvatham, V. Sushant, Bijay Kumar Rout

Abstract:

In this paper, we introduce a new computer vision task called nighttime dehaze-enhancement. This task aims to jointly perform dehazing and lightness enhancement. Our task fundamentally differs from nighttime dehazing – our goal is to jointly dehaze and enhance scenes, while nighttime dehazing aims to dehaze scenes under a nighttime setting. In order to facilitate further research on this task, we release a new benchmark dataset called Reside-β Night dataset, consisting of 4122 nighttime hazed images from 2061 scenes and 2061 ground truth images. Moreover, we also propose a new network called NDENet (Nighttime Dehaze-Enhancement Network), which jointly performs dehazing and low-light enhancement in an end-to-end manner. We evaluate our method on the proposed benchmark and achieve SSIM of 0.8962 and PSNR of 26.25. We also compare our network with other baseline networks on our benchmark to demonstrate the effectiveness of our approach. We believe that nighttime dehaze-enhancement is an essential task, particularly for autonomous navigation applications, and we hope that our work will open up new frontiers in research. Our dataset and code will be made publicly available upon acceptance of our paper.

Keywords: dehazing, image enhancement, nighttime, computer vision

Procedia PDF Downloads 118

Authors: Mo-Se Lee, Cheol-Hwi Ahn, Kee-Young Kwahk, Hyunchul Ahn

Abstract:

Over the past decade, deep learning has been in spotlight among various machine learning algorithms. In particular, CNN (Convolutional Neural Network), which is known as effective solution for recognizing and classifying images, has been popularly applied to classification and prediction problems in various fields. In this study, we try to apply CNN to stock market prediction, one of the most challenging tasks in the machine learning research. In specific, we propose to apply CNN as the binary classifier that predicts stock market direction (up or down) by using a graph as its input. That is, our proposal is to build a machine learning algorithm that mimics a person who looks at the graph and predicts whether the trend will go up or down. Our proposed model consists of four steps. In the first step, it divides the dataset into 5 days, 10 days, 15 days, and 20 days. And then, it creates graphs for each interval in step 2. In the next step, CNN classifiers are trained using the graphs generated in the previous step. In step 4, it optimizes the hyper parameters of the trained model by using the validation dataset. To validate our model, we will apply it to the prediction of KOSPI200 for 1,986 days in eight years (from 2009 to 2016). The experimental dataset will include 14 technical indicators such as CCI, Momentum, ROC and daily closing price of KOSPI200 of Korean stock market.

Keywords: convolutional neural network, deep learning, Korean stock market, stock market prediction

Procedia PDF Downloads 406

Authors: Kemal Polat

Abstract:

In this study, it has been estimated whether there is skin by using RBG values obtained from the camera and k-nearest neighbor (k-NN) classifier. The dataset used in this study has an unbalanced distribution and a linearly non-separable structure. This problem can also be called a big data problem. The Skin dataset was taken from UCI machine learning repository. As the classifier, we have used the k-NN method to handle this big data problem. For k value of k-NN classifier, we have used as 1. To train and test the k-NN classifier, 50-50% training-testing partition has been used. As the performance metrics, TP rate, FP Rate, Precision, recall, f-measure and AUC values have been used to evaluate the performance of k-NN classifier. These obtained results are as follows: 0.999, 0.001, 0.999, 0.999, 0.999, and 1,00. As can be seen from the obtained results, this proposed method could be used to predict whether the image is skin or not.

Keywords: k-NN classifier, skin or non-skin classification, RGB values, classification

Procedia PDF Downloads 220

Authors: Gurukiran Veerapur, Nytik Birudavolu, Seetharam U. N., Chandravva Hebbi, R. Praneeth Reddy

Abstract:

This work presents a robust approach for the recognition of characters in Telugu and Kannada, two South Indian scripts with structural similarities in characters. To recognize the characters exhaustive datasets are required, but there are only a few publicly available datasets. As a result, we decided to create a dataset for one language (source language),train the model with it, and then test it with the target language.Telugu is the target language in this work, whereas Kannada is the source language. The suggested method makes use of Canny edge features to increase character identification accuracy on pictures with noise and different lighting. A dataset of 45,150 images containing printed Kannada characters was created. The Nudi software was used to automatically generate printed Kannada characters with different writing styles and variations. Manual labelling was employed to ensure the accuracy of the character labels. The deep learning models like CNN (Convolutional Neural Network) and Visual Attention neural network (VAN) are used to experiment with the dataset. A Visual Attention neural network (VAN) architecture was adopted, incorporating additional channels for Canny edge features as the results obtained were good with this approach. The model's accuracy on the combined Telugu and Kannada test dataset was an outstanding 97.3%. Performance was better with Canny edge characteristics applied than with a model that solely used the original grayscale images. The accuracy of the model was found to be 80.11% for Telugu characters and 98.01% for Kannada words when it was tested with these languages. This model, which makes use of cutting-edge machine learning techniques, shows excellent accuracy when identifying and categorizing characters from these scripts.

Keywords: base characters, modifiers, guninthalu, aksharas, vattakshara, VAN

Procedia PDF Downloads 26

Authors: Shourav Chowdhury, Subrata Barua, K. M. Naimuddin, Imam Hassan Sajib, Md. Hasan, Shudipta Banik, Muna Das

Abstract:

Traffic accidents stand as a leading cause of fatalities worldwide, significantly impacting global mortality rates. Accurate classification of road accidents through advanced technological solutions presents a crucial opportunity to revolutionize accident prevention and emergency response strategies. This paper presents an advanced deep-learning methodology customized for the classification of road accidents using CCTV surveillance footage. This real-time dataset, comprising approximately 18,000 frames, has been amassed, which is pivotal for enabling comprehensive research in this field. This substantial dataset is the foundation for these investigative efforts, providing a rich and diverse source for conducting an in-depth analysis of the features. It has achieved a remarkable accuracy of 97% on this dataset through the strategic utilization of transfer learning in conjunction with LSTM (Long short-term memory) techniques. This accomplishment underscores the efficacy of our approach, combining the strengths of transfer learning and LSTM models, resulting in a highly accurate classification system for road accident events.

Keywords: accident, CCTV, footage, long short-term memory, surveillance

Procedia PDF Downloads 24

Authors: Hangsik Shin

Abstract:

The purpose of this research is to restore the feature location of under-sampled photoplethysmogram using spline interpolation and to investigate feasibility for feature shape restoration. We obtained 10 kHz-sampled photoplethysmogram and decimated it to generate under-sampled dataset. Decimated dataset has 5 kHz, 2.5 k Hz, 1 kHz, 500 Hz, 250 Hz, 25 Hz and 10 Hz sampling frequency. To investigate the restoration performance, we interpolated under-sampled signals with 10 kHz, then compared feature locations with feature locations of 10 kHz sampled photoplethysmogram. Features were upper and lower peak of photplethysmography waveform. Result showed that time differences were dramatically decreased by interpolation. Location error was lesser than 1 ms in both feature types. In 10 Hz sampled cases, location error was also deceased a lot, however, they were still over 10 ms.

Keywords: peak detection, photoplethysmography, sampling, signal reconstruction

Procedia PDF Downloads 338

Authors: Basma Khelfa, Antoine Tordeux, Ibrahima Ba

Abstract:

We present a descriptive analysis data of lane-changing events in multi-lane roads. The data are provided from The Highway Drone Dataset (HighD), which are microscopic trajectories in highway. This paper describes and analyses the role of the different parameters and their significance. Thanks to HighD data, we aim to find the most frequent reasons that motivate drivers to change lanes. We used the programming language R for the processing of these data. We analyze the involvement and relationship of different variables of each parameter of the ego vehicle and the four vehicles surrounding it, i.e., distance, speed difference, time gap, and acceleration. This was studied according to the class of the vehicle (car or truck), and according to the maneuver it undertook (overtaking or falling back).

Keywords: autonomous driving, physical traffic model, prediction model, statistical learning process

Procedia PDF Downloads 229

Authors: Tamanna Jethava, Rahul Joshi

Abstract:

Frequent Item sets play an essential role in many data Mining tasks that try to find interesting patterns from the database. Typically it refers to a set of items that frequently appear together in transaction dataset. There are several mining algorithm being used for frequent item set mining, yet most do not scale to the type of data we presented with today, so called “BIG DATA”. Big Data is a collection of large data sets. Our approach is to work on the frequent item set mining over the large dataset with scalable and speedy way. Big Data basically works with Map Reduce along with HDFS is used to find out frequent item sets from Big Data on large cluster. This paper focuses on using pre-processing & mining algorithm as hybrid approach for big data over Hadoop platform.

Keywords: frequent item set mining, big data, Hadoop, MapReduce

Procedia PDF Downloads 389

Authors: Shayan Mohajer Hamidi

Abstract:

Federated Learning (FL) has emerged as a promising paradigm for training machine learning models on distributed data without the need for centralized data aggregation. In the realm of wireless communications, FL has the potential to leverage the vast amounts of data generated by wireless devices to improve model performance and enable intelligent applications. However, the fairness aspect of FL in wireless communications remains largely unexplored. This abstract presents an idea for fair federated learning in wireless communications, addressing the challenges of imbalanced data distribution, privacy preservation, and resource allocation. Firstly, the proposed approach aims to tackle the issue of imbalanced data distribution in wireless networks. In typical FL scenarios, the distribution of data across wireless devices can be highly skewed, resulting in unfair model updates. To address this, we propose a weighted aggregation strategy that assigns higher importance to devices with fewer samples during the aggregation process. By incorporating fairness-aware weighting mechanisms, the proposed approach ensures that each participating device's contribution is proportional to its data distribution, thereby mitigating the impact of data imbalance on model performance. Secondly, privacy preservation is a critical concern in federated learning, especially in wireless communications where sensitive user data is involved. The proposed approach incorporates privacy-enhancing techniques, such as differential privacy, to protect user privacy during the model training process. By adding carefully calibrated noise to the gradient updates, the proposed approach ensures that the privacy of individual devices is preserved without compromising the overall model accuracy. Moreover, the approach considers the heterogeneity of devices in terms of computational capabilities and energy constraints, allowing devices to adaptively adjust the level of privacy preservation to strike a balance between privacy and utility. Thirdly, efficient resource allocation is crucial for federated learning in wireless communications, as devices operate under limited bandwidth, energy, and computational resources. The proposed approach leverages optimization techniques to allocate resources effectively among the participating devices, considering factors such as data quality, network conditions, and device capabilities. By intelligently distributing the computational load, communication bandwidth, and energy consumption, the proposed approach minimizes resource wastage and ensures a fair and efficient FL process in wireless networks. To evaluate the performance of the proposed fair federated learning approach, extensive simulations and experiments will be conducted. The experiments will involve a diverse set of wireless devices, ranging from smartphones to Internet of Things (IoT) devices, operating in various scenarios with different data distributions and network conditions. The evaluation metrics will include model accuracy, fairness measures, privacy preservation, and resource utilization. The expected outcomes of this research include improved model performance, fair allocation of resources, enhanced privacy preservation, and a better understanding of the challenges and solutions for fair federated learning in wireless communications. The proposed approach has the potential to revolutionize wireless communication systems by enabling intelligent applications while addressing fairness concerns and preserving user privacy.

Keywords: federated learning, wireless communications, fairness, imbalanced data, privacy preservation, resource allocation, differential privacy, optimization

Procedia PDF Downloads 46

Authors: Oluwaponmile David Alao

Abstract:

Misdiagnosis has been the major problem in health sector. Heart attack has been one of diseases that have high level of misdiagnosis recorded on the part of physicians. In this paper, an intelligent system has been developed for diagnosis of heart attack in the health sector. Dataset of heart attack obtained from UCI repository has been used. This dataset is made up of thirteen attributes which are very vital in diagnosis of heart disease. The system is developed on the multilayer perceptron trained with back propagation neural network then simulated with feed forward neural network and a recognition rate of 87% was obtained which is a good result for diagnosis of heart attack in medical field.

Keywords: heart attack, artificial neural network, diagnosis, intelligent system

Procedia PDF Downloads 621

Authors: Tetsuro Saeki, Yuichi Kato, Shoutarou Mizuno

Abstract:

STRIM (Statistical Test Rule Induction Method) has been proposed as a method to effectively induct if-then rules from the decision table which is considered as a sample set obtained from the population of interest. Its usefulness has been confirmed by simulation experiments specifying rules in advance, and by comparison with conventional methods. However, scope for future development remains before STRIM can be applied to the analysis of real-world data sets. The first requirement is to determine the size of the dataset needed for inducting true rules, since finding statistically significant rules is the core of the method. The second is to examine the capacity of rule induction from datasets with contaminated attribute values created by missing data and noise, since real-world datasets usually contain such contaminated data. This paper examines the first problem theoretically, in connection with the rule length. The second problem is then examined in a simulation experiment, utilizing the critical size of dataset derived from the first step. The experimental results show that STRIM is highly robust in the analysis of datasets with contaminated attribute values, and hence is applicable to realworld data.

Keywords: rule induction, decision table, missing data, noise

Procedia PDF Downloads 366

Authors: Vo Ngoc Phu, Phan Thi Tuoi

Abstract:

We have explored different methods of improving the accuracy of sentiment classification. The sentiment orientation of a document can be positive (+), negative (-), or neutral (0). We combine five dictionaries from [2, 3, 4, 5, 6] into the new one with 21137 entries. The new dictionary has many verbs, adverbs, phrases and idioms, that are not in five ones before. The paper shows that our proposed method based on the combination of Term-Counting method and Enhanced Contextual Valence Shifters method has improved the accuracy of sentiment classification. The combined method has accuracy 68.984% on the testing dataset, and 69.224% on the training dataset. All of these methods are implemented to classify the reviews based on our new dictionary and the Internet Movie data set.

Keywords: sentiment classification, sentiment orientation, valence shifters, contextual, valence shifters, term counting

Procedia PDF Downloads 475

Authors: Rosa Figueroa, Christopher Flores

Abstract:

Text categorization is the problem of assigning a new document to a set of predetermined categories, on the basis of a training set of free-text data that contains documents whose category membership is known. To train a classification model, it is necessary to extract characteristics in the form of tokens that facilitate the learning and classification process. In text categorization, the feature extraction process involves the use of word sequences also known as N-grams. In general, it is expected that documents belonging to the same category share similar features. The Smith-Waterman (SW) algorithm is a dynamic programming algorithm that performs a local sequence alignment in order to determine similar regions between two strings or protein sequences. This work explores the use of SW algorithm as an alternative to feature extraction in text categorization. The dataset used for this purpose, contains 2,610 annotated documents with the classes Obese/Non-Obese. This dataset was represented in a matrix form using the Bag of Word approach. The score selected to represent the occurrence of the tokens in each document was the term frequency-inverse document frequency (TF-IDF). In order to extract features for classification, four experiments were conducted: the first experiment used SW to extract features, the second one used unigrams (single word), the third one used bigrams (two word sequence) and the last experiment used a combination of unigrams and bigrams to extract features for classification. To test the effectiveness of the extracted feature set for the four experiments, a Support Vector Machine (SVM) classifier was tuned using 20% of the dataset. The remaining 80% of the dataset together with 5-Fold Cross Validation were used to evaluate and compare the performance of the four experiments of feature extraction. Results from the tuning process suggest that SW performs better than the N-gram based feature extraction. These results were confirmed by using the remaining 80% of the dataset, where SW performed the best (accuracy = 97.10%, weighted average F-measure = 97.07%). The second best was obtained by the combination of unigrams-bigrams (accuracy = 96.04, weighted average F-measure = 95.97) closely followed by the bigrams (accuracy = 94.56%, weighted average F-measure = 94.46%) and finally unigrams (accuracy = 92.96%, weighted average F-measure = 92.90%).

Keywords: comorbidities, machine learning, obesity, Smith-Waterman algorithm

Procedia PDF Downloads 269