• The Journal of the Korea institute of electronic communication sciences
• /
• v.6 no.1
• /
• pp.34-42
• /
• 2011

In an effort to the practical realization of phonetic typewriter, we concentrate on the classification of consonants in this paper. Since many of consonants do not show periodic behavior in time domain and thus the validity for Fourier analysis of them are not convincing, vector quantization (VQ) via LBG clustering is first performed to check if the feature vectors of MFCC and LPCC are ever meaningful for consonants. Experimental results of VQ showed that it's not easy to draw a clear-cut conclusion as to the validity of Fourier analysis for consonants. For classification purpose, two kinds of neural networks are employed in our study: self organizing map (SOM) and learning vector quantization (LVQ). Results from SOM revealed that some pairs of phonemes are not resolved. Though LVQ is free from this difficulty inherently, the classification accuracy was found to be low. This suggests that, as long as consonant classification by LVQ is concerned, other types of feature vectors than MFCC should be deployed in parallel. However, the combination of MFCC/LVQ was not found to be inferior to the classification of phonemes by language-moded based approach. In all of our work, LPCC worked worse than MFCC.

• Journal of Korea Multimedia Society
• /
• v.15 no.8
• /
• pp.963-976
• /
• 2012

'Gesture' except for voice is the most intuitive means of communication. Thus, many researches on how to control computer using gesture are in progress. User detection and tracking in these studies is one of the most important processes. Conventional 2D object detection and tracking methods are sensitive to changes in the environment or lights, and a mix of 2D and 3D information methods has the disadvantage of a lot of computational complexity. In addition, using conventional 3D information methods can not segment similar depth object. In this paper, we propose object detection and tracking method using Depth Projection Map that is the cumulative value of the depth and motion information. Simulation results show that our method is robust to changes in lighting or environment, and has faster operation speed, and can work well for detection and tracking of similar depth objects.

• The Journal of the Korea Contents Association
• /
• v.6 no.12
• /
• pp.105-112
• /
• 2006

In this study, we propose a simple preprocessing method for classification of basal cell carcinoma (BCC), which is one of the most common skin cancer. The preprocessing step consists of data clipping with a half Hanning window and dimension reduction with principal components analysis (PCA). The application of the half Hanning window deemphasizes the peak near $1650cm^{-1}$ and improves classification performance by lowering the false negative ratio. Classification results with various classifiers are presented to show the effectiveness of the proposed method. The classifiers include maximum a posteriori probability (MAP), k-nearest neighbor (KNN), probabilistic neural network (PNN), multilayer perceptron(MLP), support vector machine (SVM) and minimum squared error (MSE) classification. Classification results with KNN involving 216 spectra preprocessed with the proposed method gave 97.3% sensitivity, which is very promising results for automatic BCC detection.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.34 no.2
• /
• pp.207-215
• /
• 2016

The demand for spatial data service technologies is increasing lately with the development of realistic 3D spatial information services and ICT (Information and Communication Technology). Research is being conducted on the real-time provision of spatial data services through a variety of mobile and Web-based contents. Big data or cloud computing can be presented as alternatives to the construction of spatial data for the effective use of large volumes of data. In this paper, the process of building HLM (Hyper Live Map) using multi-source data to acquire stereo CCTV and other various data is presented and a big data service architecture design is proposed for the use of flexible and scalable cloud computing to handle big data created by users through such media as social network services and black boxes. The provision of spatial data services in real time using big data and cloud computing will enable us to implement navigation systems, vehicle augmented reality, real-time 3D spatial information, and single picture based positioning above the single GPS level using low-cost image-based position recognition technology in the future. Furthermore, Big Data and Cloud Computing are also used for data collection and provision in U-City and Smart-City environment as well, and the big data service architecture will provide users with information in real time.

• Journal of the Korea Society of Computer and Information
• /
• v.20 no.6
• /
• pp.87-97
• /
• 2015

An Aerial video summarization is not only the key to effective browsing video within a limited time, but also an embedded cue to efficiently congregative situation awareness acquired by unmanned aerial vehicle. Different with previous works, we utilize sensor operation mode of unmanned aerial vehicle, which is global, local, and focused surveillance mode in order for accurately summarizing the aerial video considering flight and surveillance/reconnaissance environments. In focused mode, we propose the moving-react tracking method which utilizes the partitioning motion vector and spatiotemporal saliency map to detect and track the interest moving object continuously. In our simulation result, the key frames are correctly detected for aerial video summarization according to the sensor operation mode of aerial vehicle and finally, we verify the efficiency of video summarization using the proposed mothed.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.12 no.5
• /
• pp.432-441
• /
• 2002

Though many services offer useful information on internet, computer users are not so familiar with such services that they need an assistant system to use the services easily In the case of web sites, for example, the operators answer the users e-mail questions, but the increasing number of users makes it hard to answer the questions efficiently. In this paper, we propose an assistant system which responds to the users questions automatically and helps them browse the Hanmail Net FAQ (Frequently Asked Question) conceptually. This system uses two-level self-organizing map (SOM): the keyword clustering SOM and document classification SOM. The keyword clustering SOM reduces a variable length question to a normalized vector and the document classification SOM classifies the question into an answer class. Experiments on the 2,206 e-mail question data collected for a month from the Hanmail net show that this system is able to find the correct answers with the recognition rate of 95% and also the browsing based on the map is conceptual and efficient.

• Journal of the Architectural Institute of Korea Structure & Construction
• /
• v.36 no.3
• /
• pp.113-120
• /
• 2020

The purpose of this study is to compare the models of Deep Learning-based Convolution Neural Network(CNN) for concrete crack detection. The comparison models are AlexNet, GoogLeNet, VGG16, VGG19, ResNet-18, ResNet-50, ResNet-101, and SqueezeNet which won ImageNet Large Scale Visual Recognition Challenge(ILSVRC). To train, validate and test these models, we constructed 3000 training data and 12000 validation data with 256×256 pixel resolution consisting of cracked and non-cracked images, and constructed 5 test data with 4160×3120 pixel resolution consisting of concrete images with crack. In order to increase the efficiency of the training, transfer learning was performed by taking the weight from the pre-trained network supported by MATLAB. From the trained network, the validation data is classified into crack image and non-crack image, yielding True Positive (TP), True Negative (TN), False Positive (FP), False Negative (FN), and 6 performance indicators, False Negative Rate (FNR), False Positive Rate (FPR), Error Rate, Recall, Precision, Accuracy were calculated. The test image was scanned twice with a sliding window of 256×256 pixel resolution to classify the cracks, resulting in a crack map. From the comparison of the performance indicators and the crack map, it was concluded that VGG16 and VGG19 were the most suitable for detecting concrete cracks.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.5
• /
• pp.55-67
• /
• 2022

Sensor data can provide fault diagnosis for equipment. However, the cause analysis for fault results of equipment is not often provided. In this study, we propose an explainable convolutional neural network framework for the sensor-based time series classification model. We used sensor-based time series dataset, acquired from vehicles equipped with sensors, and the Wafer dataset, acquired from manufacturing process. Moreover, we used Cycle Signal dataset, acquired from real world mechanical equipment, and for Data augmentation methods, scaling and jittering were used to train our deep learning models. In addition, our proposed classification models are convolutional neural network based models, FCN, 1D-CNN, and ResNet, to compare evaluations for each model. Our experimental results show that the ResNet provides promising results in the context of time series classification with accuracy and F1 Score reaching 95%, improved by 3% compared to the previous study. Furthermore, we propose XAI methods, Class Activation Map and Layer Visualization, to interpret the experiment result. XAI methods can visualize the time series interval that shows important factors for sensor data classification.

• Journal of Drive and Control
• /
• v.20 no.4
• /
• pp.1-8
• /
• 2023

In recent years, robots have been utilized in various industries to reduce workload and enhance work efficiency. The following mobility offers users convenience by autonomously tracking specific locations and targets without the need for additional equipment such as forklifts or carts. In this paper, deep learning techniques were employed to recognize individuals and assign each of them a unique identifier to enable the recognition of a specific person even among multiple individuals. To achieve this, the distance and angle between the robot and the targeted individual are transmitted to respective controllers. Furthermore, this study explored the control methodology for mobility that tracks a specific person, utilizing Simultaneous Localization and Mapping (SLAM) and Proportional-Integral-Derivative (PID) control techniques. In the PID control method, a genetic algorithm is employed to extract the optimal gain value, subsequently evaluating PID performance through simulation. The SLAM method involves generating a map by synchronizing data from a 2D LiDAR and a depth camera using Real-Time Appearance-Based Mapping (RTAB-MAP). Experiments are conducted to compare and analyze the performance of the two control methods, visualizing the paths of both the human and the following mobility.

• Journal of IKEEE
• /
• v.28 no.1
• /
• pp.65-71
• /
• 2024

This paper proposes a neural network that can reduce the number of parameters while reducing matching errors in occluded regions to increase the accuracy of depth maps in stereo matching. Stereo matching-based object recognition is utilized in many fields to more accurately recognize situations using images. When there are many objects in a complex image, an occluded area is generated due to overlap between objects and occlusion by background, thereby lowering the accuracy of the depth map. To solve this problem, existing research methods that create context information and combine it with the cost volume or RoIselect in the occluded area increase the complexity of neural networks, making it difficult to learn and expensive to implement. In this paper, we create a depthwise seperable neural network that enhances regional feature extraction before cost volume generation, reducing the number of parameters and proposing a neural network that is robust to occlusion errors. Compared to PSMNet, the proposed neural network reduced the number of parameters by 30%, improving 5.3% in color error and 3.6% in test loss.