• International journal of advanced smart convergence
• /
• v.10 no.1
• /
• pp.1-11
• /
• 2021

Document page segmentation is an important step in building a quality optical character recognition module. The study examined already existing work on the topic of page segmentation and focused on the development of a segmentation model that has greater functional significance for application in an organization, as well as broad capabilities for managing the quality of the model. The main problems of document segmentation were highlighted, which include a complex background of intersecting objects. As classes for detection, not only classic text, table and figure were selected, but also additional types, such as signature, logo and table without borders (or with partially missing borders). This made it possible to pose a non-trivial task of detecting non-standard document elements. The authors compared existing neural network architectures for object detection based on published research data. The most suitable architecture was RetinaNet. To ensure the possibility of quality control of the model, a method based on neural network modeling using the RetinaNet architecture is proposed. During the study, several models were built, the quality of which was assessed on the test sample using the Mean average Precision metric. The best result among the constructed algorithms was shown by a model that includes four neural networks: the focus of the first neural network on detecting tables and tables without borders, the second - seals and signatures, the third - pictures and logos, and the fourth - text. As a result of the analysis, it was revealed that the approach based on four neural networks showed the best results in accordance with the objectives of the study on the test sample in the context of most classes of detection. The method proposed in the article can be used to recognize other objects. A promising direction in which the analysis can be continued is the segmentation of tables; the areas of the table that differ in function will act as classes: heading, cell with a name, cell with data, empty cell.

• Smart Structures and Systems
• /
• v.31 no.4
• /
• pp.335-349
• /
• 2023

Bridges constantly undergo deterioration and damage, the most common ones being concrete damage and exposed rebar. Periodic inspection of bridges to identify damages can aid in their quick remediation. Likewise, identifying components can provide context for damage assessment and help gauge a bridge's state of interaction with its surroundings. Current inspection techniques rely on manual site visits, which can be time-consuming and costly. More recently, robotic inspection assisted by autonomous data analytics based on Computer Vision (CV) and Artificial Intelligence (AI) has been viewed as a suitable alternative to manual inspection because of its efficiency and accuracy. To aid research in this avenue, this study performs a comparative assessment of different architectures, loss functions, and ensembling strategies for the autonomous segmentation of bridge components and damages. The experiments lead to several interesting discoveries. Nested Reg-UNet architecture is found to outperform five other state-of-the-art architectures in both damage and component segmentation tasks. The architecture is built by combining a Nested UNet style dense configuration with a pretrained RegNet encoder. In terms of the mean Intersection over Union (mIoU) metric, the Nested Reg-UNet architecture provides an improvement of 2.86% on the damage segmentation task and 1.66% on the component segmentation task compared to the state-of-the-art UNet architecture. Furthermore, it is demonstrated that incorporating the Lovasz-Softmax loss function to counter class imbalance can boost performance by 3.44% in the component segmentation task over the most employed alternative, weighted Cross Entropy (wCE). Finally, weighted softmax ensembling is found to be quite effective when used synchronously with the Nested Reg-UNet architecture by providing mIoU improvement of 0.74% in the component segmentation task and 1.14% in the damage segmentation task over a single-architecture baseline. Overall, the best mIoU of 92.50% for the component segmentation task and 84.19% for the damage segmentation task validate the feasibility of these techniques for autonomous bridge component and damage segmentation using RGB images.

• Smart Media Journal
• /
• v.12 no.5
• /
• pp.28-35
• /
• 2023

In this study, we propose a new light-weight model RoutingConvNet with fewer parameters to improve the applicability and practicality of speech emotion recognition. To reduce the number of learnable parameters, the proposed model connects bidirectional MFCCs on a channel-by-channel basis to learn long-term emotion dependence and extract contextual features. A light-weight deep CNN is constructed for low-level feature extraction, and self-attention is used to obtain information about channel and spatial signals in speech signals. In addition, we apply dynamic routing to improve the accuracy and construct a model that is robust to feature variations. The proposed model shows parameter reduction and accuracy improvement in the overall experiments of speech emotion datasets (EMO-DB, RAVDESS, and IEMOCAP), achieving 87.86%, 83.44%, and 66.06% accuracy respectively with about 156,000 parameters. In this study, we proposed a metric to calculate the trade-off between the number of parameters and accuracy for performance evaluation against light-weight.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.7
• /
• pp.89-98
• /
• 2024

In this paper, we propose a system for recommending lecture videos at the chapter level, addressing the balance between accuracy and processing speed in chapter-level video recommendations. Specifically, it has been observed that enhancing recommendation accuracy reduces processing speed, while increasing processing speed decreases accuracy. To mitigate this trade-off, a hybrid approach is proposed, utilizing techniques such as TF-IDF, k-means++ clustering, and Graph Neural Networks (GNN). The approach involves pre-constructing clusters based on chapter similarity to reduce computational load during recommendations, thereby improving processing speed, and applying GNN to the graph of clusters as nodes to enhance recommendation accuracy. Experimental results indicate that the use of GNN resulted in an approximate 19.7% increase in recommendation accuracy, as measured by the Mean Reciprocal Rank (MRR) metric, and an approximate 27.7% increase in precision defined by similarities. These findings are expected to contribute to the development of a learning system that recommends more suitable video chapters in response to learners' queries.

• Journal of the Korea Society of Computer and Information
• /
• v.25 no.11
• /
• pp.27-40
• /
• 2020

In this paper, we investigate the correlation between the amount of English sentences that Korean English learners (L2ers) are exposed to and their sentence processing patterns by examining what Long Short-Term Memory (LSTM) language models (LMs) can learn about implicit syntactic relationship: that is, the filler-gap dependency. The filler-gap dependency refers to a relationship between a (wh-)filler, which is a wh-phrase like 'what' or 'who' overtly in clause-peripheral position, and its gap in clause-internal position, which is an invisible, empty syntactic position to be filled by the (wh-)filler for proper interpretation. Here to implement L2ers' English learning, we build LSTM LMs that in turn learn a subset of the known restrictions on the filler-gap dependency from English sentences in the L2 corpus that L2ers can potentially encounter in their English learning. Examining LSTM LMs' behaviors on controlled sentences designed with the filler-gap dependency, we show the characteristics of L2ers' sentence processing using the information-theoretic metric of surprisal that quantifies violations of the filler-gap dependency or wh-licensing interaction effects. Furthermore, comparing L2ers' LMs with native speakers' LM in light of processing the filler-gap dependency, we not only note that in their sentence processing both L2ers' LM and native speakers' LM can track abstract syntactic structures involved in the filler-gap dependency, but also show using linear mixed-effects regression models that there exist significant differences between them in processing such a dependency.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.12
• /
• pp.53-59
• /
• 2021

In this paper, we propose a BERGPT-chatbot, a domestic AI chatbot that can alleviate negative emotions based on text input such as 'Replika'. We made BERGPT-chatbot into a chatbot capable of mitigating negative emotions by pipelined two models, KR-BERT and KoGPT2-chatbot. We applied a creative method of giving emotions to unrefined everyday datasets through KR-BERT, and learning additional datasets through KoGPT2-chatbot. The development background of BERGPT-chatbot is as follows. Currently, the number of people with depression is increasing all over the world. This phenomenon is emerging as a more serious problem due to COVID-19, which causes people to increase long-term indoor living or limit interpersonal relationships. Overseas artificial intelligence chatbots aimed at relieving negative emotions or taking care of mental health care, have increased in use due to the pandemic. In Korea, Psychological diagnosis chatbots similar to those of overseas cases are being operated. However, as the domestic chatbot is a system that outputs a button-based answer rather than a text input-based answer, when compared to overseas chatbots, domestic chatbots remain at a low level of diagnosing human psychology. Therefore, we proposed a chatbot that helps mitigating negative emotions through BERGPT-chatbot. Finally, we compared BERGPT-chatbot and KoGPT2-chatbot through 'Perplexity', an internal evaluation metric for evaluating language models, and showed the superity of BERGPT-chatbot.

• Korean Journal of Remote Sensing
• /
• v.39 no.2
• /
• pp.143-156
• /
• 2023

To precisely and reliably analyze the contents of the satellite imagery, recognizing the clouds which are the obstacle to gathering the useful information is essential. In recent times, deep learning yielded satisfactory results in various tasks, so many studies using deep neural networks have been conducted to improve the performance of cloud detection. However, existing methods for cloud detection have the limitation on increasing the performance due to the adopting the network models for semantic image segmentation without modification. To tackle this problem, we introduced the multi-branch neural architecture search to find optimal network structure for cloud detection. Additionally, the proposed method adopts the soft intersection over union (IoU) as loss function to mitigate the disagreement between the loss function and the evaluation metric and uses the various data augmentation methods. The experiments are conducted using the cloud detection dataset acquired by Arirang-3/3A satellite imagery. The experimental results showed that the proposed network which are searched network architecture using cloud dataset is 4% higher than the existing network model which are searched network structure using urban street scenes with regard to the IoU. Also, the experimental results showed that the soft IoU exhibits the best performance on cloud detection among the various loss functions. When comparing the proposed method with the state-of-the-art (SOTA) models in the field of semantic segmentation, the proposed method showed better performance than the SOTA models with regard to the mean IoU and overall accuracy.

• Journal of Intelligence and Information Systems
• /
• v.23 no.3
• /
• pp.119-138
• /
• 2017

Recently, SNS has become an important channel for marketing as well as personal communication. However, cybercrime has also evolved with the development of information and communication technology, and illegal advertising is distributed to SNS in large quantity. As a result, personal information is lost and even monetary damages occur more frequently. In this study, we propose a method to analyze which sentences and documents, which have been sent to the SNS, are related to financial fraud. First of all, as a conceptual framework, we developed a matrix of conceptual characteristics of cybercriminality on SNS and emergency management. We also suggested emergency management process which consists of Pre-Cybercriminality (e.g. risk identification) and Post-Cybercriminality steps. Among those we focused on risk identification in this paper. The main process consists of data collection, preprocessing and analysis. First, we selected two words 'daechul(loan)' and 'sachae(private loan)' as seed words and collected data with this word from SNS such as twitter. The collected data are given to the two researchers to decide whether they are related to the cybercriminality, particularly financial fraud, or not. Then we selected some of them as keywords if the vocabularies are related to the nominals and symbols. With the selected keywords, we searched and collected data from web materials such as twitter, news, blog, and more than 820,000 articles collected. The collected articles were refined through preprocessing and made into learning data. The preprocessing process is divided into performing morphological analysis step, removing stop words step, and selecting valid part-of-speech step. In the morphological analysis step, a complex sentence is transformed into some morpheme units to enable mechanical analysis. In the removing stop words step, non-lexical elements such as numbers, punctuation marks, and double spaces are removed from the text. In the step of selecting valid part-of-speech, only two kinds of nouns and symbols are considered. Since nouns could refer to things, the intent of message is expressed better than the other part-of-speech. Moreover, the more illegal the text is, the more frequently symbols are used. The selected data is given 'legal' or 'illegal'. To make the selected data as learning data through the preprocessing process, it is necessary to classify whether each data is legitimate or not. The processed data is then converted into Corpus type and Document-Term Matrix. Finally, the two types of 'legal' and 'illegal' files were mixed and randomly divided into learning data set and test data set. In this study, we set the learning data as 70% and the test data as 30%. SVM was used as the discrimination algorithm. Since SVM requires gamma and cost values as the main parameters, we set gamma as 0.5 and cost as 10, based on the optimal value function. The cost is set higher than general cases. To show the feasibility of the idea proposed in this paper, we compared the proposed method with MLE (Maximum Likelihood Estimation), Term Frequency, and Collective Intelligence method. Overall accuracy and was used as the metric. As a result, the overall accuracy of the proposed method was 92.41% of illegal loan advertisement and 77.75% of illegal visit sales, which is apparently superior to that of the Term Frequency, MLE, etc. Hence, the result suggests that the proposed method is valid and usable practically. In this paper, we propose a framework for crisis management caused by abnormalities of unstructured data sources such as SNS. We hope this study will contribute to the academia by identifying what to consider when applying the SVM-like discrimination algorithm to text analysis. Moreover, the study will also contribute to the practitioners in the field of brand management and opinion mining.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.40 no.2
• /
• pp.334-348
• /
• 2015

Widespread use of smart devices accompanies increase of use of access point (AP), which enables the connection to the wireless network. If the appropriate security is not served when a user tries to connect the wireless network through an AP, various security problems can arise due to the rogue APs. In this paper, we are going to examine the threat by evil-twin, which is a kind of rogue APs. Most of recent researches for detecting rogue APs utilize the measured time difference, such as round trip time (RTT), between the evil-twin and authorized APs. These methods, however, suffer from the low detection rate in the network congestion. Due to these reasons, in this paper, we suggest a new factor, packet inter-arrival time (PIAT), in order to detect evil-twins. By using both RTT and PIAT as the learning factors for the support vector machine (SVM), we determine the non-linear metric to classify evil-twins and authorized APs. As a result, we can detect evil-twins with the probability of up to 96.5% and at least 89.75% even when the network is congested.

• Asian-Australasian Journal of Animal Sciences
• /
• v.32 no.7
• /
• pp.1015-1026
• /
• 2019

Objective: The objective of this study was to evaluate three different degrees of white striping (WS) addressing their automatic assessment and customer acceptance. The WS classification was performed based on a computer vision system (CVS), exploring different machine learning (ML) algorithms and the most important image features. Moreover, it was verified by consumer acceptance and purchase intent. Methods: The samples for image analysis were classified by trained specialists, according to severity degrees regarding visual and firmness aspects. Samples were obtained with a digital camera, and 25 features were extracted from these images. ML algorithms were applied aiming to induce a model capable of classifying the samples into three severity degrees. In addition, two sensory analyses were performed: 75 samples properly grilled were used for the first sensory test, and 9 photos for the second. All tests were performed using a 10-cm hybrid hedonic scale (acceptance test) and a 5-point scale (purchase intention). Results: The information gain metric ranked 13 attributes. However, just one type of image feature was not enough to describe the phenomenon. The classification models support vector machine, fuzzy-W, and random forest showed the best results with similar general accuracy (86.4%). The worst performance was obtained by multilayer perceptron (70.9%) with the high error rate in normal (NORM) sample predictions. The sensory analysis of acceptance verified that WS myopathy negatively affects the texture of the broiler breast fillets when grilled and the appearance attribute of the raw samples, which influenced the purchase intention scores of raw samples. Conclusion: The proposed system has proved to be adequate (fast and accurate) for the classification of WS samples. The sensory analysis of acceptance showed that WS myopathy negatively affects the tenderness of the broiler breast fillets when grilled, while the appearance attribute of the raw samples eventually influenced purchase intentions.