• Journal of Broadcast Engineering
• /
• v.27 no.1
• /
• pp.20-30
• /
• 2022

Recently, as VR (Virtual Reality) technology has been in the spotlight, 360° panoramic images that can view lively VR contents are attracting a lot of attention. Image stitching technology is a major technology for producing 360° panorama images, and many studies are being actively conducted. Typical stitching algorithms are based on feature point-based image stitching. However, conventional feature point-based image stitching methods have a problem that stitching results are intensely affected by feature points. To solve this problem, deep learning-based image stitching technologies have recently been studied, but there are still many problems when there are few overlapping areas between images or large parallax. In addition, there is a limit to complete supervised learning because labeled ground-truth panorama images cannot be obtained in a real environment. Therefore, we produced three fisheye images with different camera centers and corresponding ground truth image through carla simulator that is widely used in the autonomous driving field. We propose image stitching model that creates a 360° panorama image with the produced fisheye image. The final experimental results are virtual datasets configured similar to the actual environment, verifying stitching results that are strong against various environments and large parallax.

• Journal of the Computational Structural Engineering Institute of Korea
• /
• v.36 no.5
• /
• pp.315-321
• /
• 2023

Long-span bridges are flexible structures with low natural frequencies and damping ratios, making them susceptible to vibrational serviceability problems. However, the current design guideline of South Korea assumes a uniform threshold of wind speed or vibrational amplitude to assess the occurrence of harmful vibrations, potentially overlooking the complex vibrational patterns observed in long-span bridges. In this study, we propose a pointwise vortex-induced vibration (VIV) detection method using a deep-learning-based signalsegmentation model. Departing from conventional supervised methods of data acquisition and manual labeling, we synthesize training data by generating sinusoidal waves with an envelope to accurately represent VIV. A Fourier synchrosqueezed transform is leveraged to extract time-frequency features, which serve as input data for training a bidirectional long short-term memory model. The effectiveness of the model trained on synthetic VIV data is demonstrated through a comparison with its counterpart trained on manually labeled real datasets from an actual cable-supported bridge.

• Journal of Korean Society of Industrial and Systems Engineering
• /
• v.47 no.2
• /
• pp.1-9
• /
• 2024

New motor development requires high-speed load testing using dynamo equipment to calculate the efficiency of the motor. Abnormal noise and vibration may occur in the test equipment rotating at high speed due to misalignment of the connecting shaft or looseness of the fixation, which may lead to safety accidents. In this study, three single-axis vibration sensors for X, Y, and Z axes were attached on the surface of the test motor to measure the vibration value of vibration. Analog data collected from these sensors was used in classification models for anomaly detection. Since the classification accuracy was around only 93%, commonly used hyperparameter optimization techniques such as Grid search, Random search, and Bayesian Optimization were applied to increase accuracy. In addition, Response Surface Method based on Design of Experiment was also used for hyperparameter optimization. However, it was found that there were limits to improving accuracy with these methods. The reason is that the sampling data from an analog signal does not reflect the patterns hidden in the signal. Therefore, in order to find pattern information of the sampling data, we obtained descriptive statistics such as mean, variance, skewness, kurtosis, and percentiles of the analog data, and applied them to the classification models. Classification models using descriptive statistics showed excellent performance improvement. The developed model can be used as a monitoring system that detects abnormal conditions of the motor test.

• Journal of KIISE
• /
• v.42 no.6
• /
• pp.817-822
• /
• 2015

Relation extraction is an important information extraction technique that can be widely used in areas such as question-answering and knowledge population. Previous studies on relation extraction have been based on supervised machine learning models that need a large amount of training data manually annotated with relation categories. Recently, to reduce the manual annotation efforts for constructing training data, distant supervision methods have been proposed. However, these methods suffer from a drawback: it is difficult to use these methods for collecting negative training data that are necessary for resolving classification problems. To overcome this drawback, we propose a one-class classification model that can be trained without using negative data. The proposed model determines whether an input data item is included in an inner category by using a similarity measure based on lexical information and syntactic patterns in a vector space. In the experiments conducted in this study, the proposed model showed higher performance (an F1-score of 0.6509 and an accuracy of 0.6833) than a representative one-class classification model, one-class SVM(Support Vector Machine).

• Analytical Science and Technology
• /
• v.33 no.2
• /
• pp.98-107
• /
• 2020

Methamphetamine (MA) is currently the most abused illicit drug in Korea. MA is produced by chemical synthesis, and the final target drug that is produced contains small amounts of the precursor chemicals, intermediates, and by-products. To identify and quantify these trace compounds in MA seizures, a practical and feasible approach for conducting chromatographic fingerprinting with a suite of traditional chemometric methods and recently introduced machine learning approaches was examined. This was achieved using gas chromatography (GC) coupled with a flame ionization detector (FID) and mass spectrometry (MS). Following appropriate examination of all the peaks in 71 samples, 166 impurities were selected as the characteristic components. Unsupervised (principal component analysis (PCA), hierarchical cluster analysis (HCA), and K-means clustering) and supervised (partial least squares-discriminant analysis (PLS-DA), orthogonal partial least squares-discriminant analysis (OPLS-DA), support vector machines (SVM), and deep neural network (DNN) with Keras) chemometric techniques were employed for classifying the 71 MA seizures. The results of the PCA, HCA, K-means clustering, PLS-DA, OPLS-DA, SVM, and DNN methods for quality evaluation were in good agreement. However, the tested MA seizures possessed distinct features, such as chirality, cutting agents, and boiling points. The study indicated that the established qualitative and semi-quantitative methods will be practical and useful analytical tools for characterizing trace compounds in illicit MA seizures. Moreover, they will provide a statistical basis for identifying the synthesis route, sources of supply, trafficking routes, and connections between seizures, which will support drug law enforcement agencies in their effort to eliminate organized MA crime.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.11 no.5
• /
• pp.379-387
• /
• 2001

We proposed an enhanced extraction method of vehicle plate, in which both the brightness variation of gray and the Hue value of HSI color model were used. For the extraction of the vehicle plate from a vehicle image, first of all, candidate regions for the vehicle plate were extracted from the image by using the property of brightness variation of the image. A real place region was determined among candidate regions by the density of pixels with the Hue value of green and white. For- extracting the feature area containing characters from the extracted vehicle plate, we used the histogram-based approach of individual characters. And we proposed and applied for the recognition of characters the enhanced ART2 algorithm which support the dynamical establishment of the vigilance threshold with the genera]iced union operator of Yager. In addition, we propose an enhanced SOSL algorithm which is integrated both enhanced ART2 and supervised learning methods. The performance evaluation was performed using 100's real vehicle images and the evaluation results demonstrated that the extraction rates of tole proposed extraction method were improved, compared with that of previous methods based un brightness variation, RGB and HSI individually . Furthermore, the recognition rates of the proposed algorithms were improved much more than that of the conventional ART2 and BP algorithms.

• Research in Mathematical Education
• /
• v.23 no.4
• /
• pp.235-254
• /
• 2020

Through this study, we aimed to capture the nature of a mathematics method course, called "the Curriculum Development and Teaching Methods in Mathematics Education" which is a pedagogy course for teaching for secondary school mathematics taught at a university located in a south eastern part of South Korea. The research participants include three junior students who took the methods course and a local high school math teacher with two professors. The research has three parts. First, we designed a method course to prepare the junior or senior students for a teaching practicum. The individual students gave a mini lecture about a secondary mathematical topic as a course requirement. Second, the three students watched a classroom video-clip of the high school teacher and analyzed his instruction before the actual classroom visits. Furthermore, by "Let's Learn" program for students, the course was associated with a local community through the students and so that they could visit the teacher's classroom three times to observe his math classroom teaching. The students discussed the difference between their own mini lectures and the actual math classroom teaching to develop an understanding of what it entails to teach an actual math class. Third, the first author supervised the students' activities in the program including their report for it to bring out their findings to the class of the method course. We found out this method course provided the students with the experience of various aspects of actual math lesson as well as learning theories about the pedagogy for teaching for secondary school mathematics. We conclude that this course gives a model for the method course in mathematics education for secondary school mathematics.

• Journal of the Korea Computer Graphics Society
• /
• v.22 no.3
• /
• pp.21-29
• /
• 2016

Cell segmentation is an important but time-consuming and laborious task in biological image analysis. An automated, robust, and fast method is required to overcome such burdensome processes. These needs are, however, challenging due to various cell shapes, intensity, and incomplete boundaries. A precise cell segmentation will allow to making a pathological diagnosis of tissue samples. A vast body of literature exists on cell segmentation in microscopy images [1]. The majority of existing work is based on input images and predefined feature models only - for example, using a deformable model to extract edge boundaries in the image. Only a handful of recent methods employ data-driven approaches, such as supervised learning. In this paper, we propose a novel data-driven cell segmentation algorithm for bright-field microscopy images. The proposed method minimizes an energy formula defined by two dictionaries - one is for input images and the other is for their manual segmentation results - and a common sparse code, which aims to find the pixel-level classification by deploying the learned dictionaries on new images. In contrast to deformable models, we do not need to know a prior knowledge of objects. We also employed convolutional sparse coding and Alternating Direction of Multiplier Method (ADMM) for fast dictionary learning and energy minimization. Unlike an existing method [1], our method trains both dictionaries concurrently, and is implemented using the GPU device for faster performance.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.75-82
• /
• 2023

Automatic speech recognition (ASR) has been revolutionized with deep learning-based approaches, among which self-supervised learning methods have proven to be particularly effective. In this study, we aim to enhance the performance of OpenAI's Whisper model, a multilingual ASR system on the Korean language. Whisper was pretrained on a large corpus (around 680,000 hours) of web speech data and has demonstrated strong recognition performance for major languages. However, it faces challenges in recognizing languages such as Korean, which is not major language while training. We address this issue by fine-tuning the Whisper model with an additional dataset comprising about 1,000 hours of Korean speech. We also compare its performance against a Transformer model that was trained from scratch using the same dataset. Our results indicate that fine-tuning the Whisper model significantly improved its Korean speech recognition capabilities in terms of character error rate (CER). Specifically, the performance improved with increasing model size. However, the Whisper model's performance on English deteriorated post fine-tuning, emphasizing the need for further research to develop robust multilingual models. Our study demonstrates the potential of utilizing a fine-tuned Whisper model for Korean ASR applications. Future work will focus on multilingual recognition and optimization for real-time inference.

• Journal of Gastric Cancer
• /
• v.22 no.2
• /
• pp.120-134
• /
• 2022

Purpose: This study aimed to identify prognostic factors for patients with distant lymph node-involved gastric cancer (GC) using a machine learning algorithm, a method that offers considerable advantages and new prospects for high-dimensional biomedical data exploration. Materials and Methods: This study employed 79 features of clinical pathology, laboratory tests, and therapeutic details from 289 GC patients whose distant lymphadenopathy was presented as the first episode of recurrence or metastasis. Outcomes were measured as any-cause death events and survival months after distant lymph node metastasis. A prediction model was built based on possible outcome predictors using a random survival forest algorithm and confirmed by 5×5 nested cross-validation. The effects of single variables were interpreted using partial dependence plots. A contour plot was used to visually represent survival prediction based on 2 predictive features. Results: The median survival time of patients with GC with distant nodal metastasis was 9.2 months. The optimal model incorporated the prealbumin level and the prothrombin time (PT), and yielded a prediction error of 0.353. The inclusion of other variables resulted in poorer model performance. Patients with higher serum prealbumin levels or shorter PTs had a significantly better prognosis. The predicted one-year survival rate was stratified and illustrated as a contour plot based on the combined effect the prealbumin level and the PT. Conclusions: Machine learning is useful for identifying the important determinants of cancer survival using high-dimensional datasets. The prealbumin level and the PT on distant lymph node metastasis are the 2 most crucial factors in predicting the subsequent survival time of advanced GC.