• Korean Journal of Remote Sensing
• /
• v.35 no.2
• /
• pp.279-288
• /
• 2019

We evaluated the land cover classification performance of SegNet, which features semantic segmentation of aerial imagery. We selected four semantic classes, i.e., urban, farmland, forest, and water areas, and created 2,000 datasets using aerial images and land cover maps. The datasets were divided at a 8:2 ratio into training (1,600) and validation datasets (400); we evaluated validation accuracy after tuning the hyperparameters. SegNet performance was optimal at a batch size of five with 100,000 iterations. When 200 test datasets were subjected to semantic segmentation using the trained SegNet model, the accuracies were farmland 87.89%, forest 87.18%, water 83.66%, and urban regions 82.67%; the overall accuracy was 85.48%. Thus, deep learning-based semantic segmentation can be used to classify land cover.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.4
• /
• pp.29-37
• /
• 2021

In this paper, we propose a deep learning-based retinal vessel segmentation model for handling multi-scale information of fundus images. we integrate the selective kernel convolution into U-Net-based convolutional neural network. The proposed model extracts and segment features information with various shapes and sizes of retinal blood vessels, which is important information for diagnosing eye-related diseases from fundus images. The proposed model consists of standard convolutions and selective kernel convolutions. While the standard convolutional layer extracts information through the same size kernel size, The selective kernel convolution extracts information from branches with various kernel sizes and combines them by adaptively adjusting them through split-attention. To evaluate the performance of the proposed model, we used the DRIVE and CHASE DB1 datasets and the proposed model showed F1 score of 82.91% and 81.71% on both datasets respectively, confirming that the proposed model is effective in segmenting retinal blood vessels.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.2
• /
• pp.65-72
• /
• 2021

Recently, attacks on the network environment have been rapidly escalating and intelligent. Thus, the signature-based network intrusion detection system is becoming clear about its limitations. To solve these problems, research on machine learning-based intrusion detection systems is being conducted in many ways, but two problems are encountered to use machine learning for intrusion detection. The first is to find important features associated with learning for real-time detection, and the second is the imbalance of data used in learning. This problem is fatal because the performance of machine learning algorithms is data-dependent. In this paper, we propose the HSF-DNN, a network intrusion detection model based on a deep neural network to solve the problems presented above. The proposed HFS-DNN was learned through the NSL-KDD data set and performs performance comparisons with existing classification models. Experiments have confirmed that the proposed Hybrid Feature Selection algorithm does not degrade performance, and in an experiment between learning models that solved the imbalance problem, the model proposed in this paper showed the best performance.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.38 no.6
• /
• pp.635-644
• /
• 2020

Object recognition, detection and instance segmentation based on DL (Deep Learning) have being used in various practices, and mainly optical images are used as training data for DL models. The major objective of this paper is object segmentation and building detection by utilizing multimodal datasets as well as optical images for training Detectron2 model that is one of the improved R-CNN (Region-based Convolutional Neural Network). For the implementation, infrared aerial images, LiDAR data, and edges from the images, and Haralick features, that are representing statistical texture information, from LiDAR (Light Detection And Ranging) data were generated. The performance of the DL models depends on not only on the amount and characteristics of the training data, but also on the fusion method especially for the multimodal data. The results of segmenting objects and detecting buildings by applying hybrid fusion - which is a mixed method of early fusion and late fusion - results in a 32.65% improvement in building detection rate compared to training by optical image only. The experiments demonstrated complementary effect of the training multimodal data having unique characteristics and fusion strategy.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.15 no.6
• /
• pp.987-994
• /
• 2020

In this paper, we studied the effect of kernel size of CNN layer on performance in acoustic noise attenuators. This system uses a deep learning algorithm using a neural network adaptive prediction filter instead of using the existing adaptive filter. Speech is estimated from a single input speech signal containing noise using a 100-neuron, 16-filter CNN filter and an error back propagation algorithm. This is to use the quasi-periodic property in the voiced sound section of the voice signal. In this study, a simulation program using Tensorflow and Keras libraries was written and a simulation was performed to verify the performance of the noise attenuator for the kernel size. As a result of the simulation, when the kernel size is about 16, the MSE and MAE values are the smallest, and when the size is smaller or larger than 16, the MSE and MAE values increase. It can be seen that in the case of an speech signal, the features can be best captured when the kernel size is about 16.

• Investigative Magnetic Resonance Imaging
• /
• v.25 no.2
• /
• pp.101-108
• /
• 2021

Purpose: To identify characteristic magnetic resonance imaging (MRI) features to differentiate between Krabbe disease and metachromatic leukodystrophy (MLD) in young children. Materials and Methods: We collected all confirmed cases of Krabbe disease and MLD between October 2004 and September 2020 at Seoul National University Children's Hospital. Patients with initial MRI available were included. Their initial MRIs were retrospectively reviewed for the following: 1) presence of white matter signal abnormality involving the periventricular and deep white matter, subcortical white matter, internal capsule, brainstem, and cerebellum; 2) presence of volume decrease and signal alteration in the corpus callosum and thalamus; 3) presence of the tigroid sign; 4) presence of optic nerve hypertrophy; and 5) presence of enhancement or diffusion restriction. Results: Eleven children with Krabbe disease and 12 children with MLD were included in this study. There was no significant difference in age or symptoms at onset. Periventricular and deep white matter signal alterations sparing the subcortical white matter were present in almost all patients of the two groups. More patients with Krabbe disease had T2 hyperintensities in the internal capsule and brainstem than patients with MLDs. In contrast, more patients with MLD had T2 hyperintensities in the splenium and genu of the corpus callosum. No patient with Krabbe disease showed T2 hyperintensity in the corpus callosal genu. A decrease in volume in the corpus callosum and thalamus was more frequently observed in patients with Krabbe disease than in those with MLD. Other MRI findings including the tigroid sign and optic nerve hypertrophy were not significantly different between the two groups. Conclusion: Signal abnormalities in the internal capsule and brainstem, decreased thalamic volume, decreased splenial volume accompanied by signal changes, and absence of signal changes in the callosal genu portion were MRI findings suggestive of Krabbe disease rather than MLD based on initial MRI. Other MRI findings such as the tigroid sign could not help differentiate between these two diseases.

• Smart Media Journal
• /
• v.9 no.4
• /
• pp.162-169
• /
• 2020

Design is a key factor that determines the competitiveness of products in the textile and fashion industry. It is very important to measure the similarity of the proposed design in order to prevent unauthorized copying and to confirm the originality. In this study, a deep learning technique was used to quantify features from images of textile designs, and similarity was measured using Spearman correlation coefficients. To verify that similar samples were actually detected, 300 images were randomly rotated and color changed. The results of Top-3 and Top-5 in the order of similarity value were measured to see if samples that rotated or changed color were detected. As a result, the VGG-16 model recorded significantly higher performance than did AlexNet. The performance of the VGG-16 model was the highest at 64% and 73.67% in the Top-3 and Top-5, where similarity results were high in the case of the rotated image. appear. In the case of color change, the highest in Top-3 and Top-5 at 86.33% and 90%, respectively.

• Journal of the Korea Society for Simulation
• /
• v.30 no.1
• /
• pp.75-85
• /
• 2021

In this paper, we present a deep neural network-based prediction model that processes and analyzes the corporate credit and personal credit information of individual business owners as a new method to predict the default rate of individual business more accurately. In modeling research in various fields, feature selection techniques have been actively studied as a method for improving performance, especially in predictive models including many features. In this paper, after statistical verification of macroeconomic indicators (macro variables) and credit information (micro variables), which are input variables used in the default rate prediction model, additionally, through the credit information feature selection method, the final feature set that improves prediction performance was identified. The proposed credit information feature selection method as an iterative & hybrid method that combines the filter-based and wrapper-based method builds submodels, constructs subsets by extracting important variables of the maximum performance submodels, and determines the final feature set through prediction performance analysis of the subset and the subset combined set.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.1
• /
• pp.76-84
• /
• 2022

Arrhythmia is a condition in which the heart beats abnormally or irregularly, early detection is very important because it can cause dangerous situations such as fainting or sudden cardiac death. However, performance degradation occurs due to personalized differences in ECG signals. In this paper, we propose arrhythmia classification using hybrid combination model of CNN-LSTM. For this purpose, the R wave is detected from noise removed signal and a single bit segment was extracted. It consisted of eight convolutional layers to extract the features of the arrhythmia in detail, used them as the input of the LSTM. The weights were learned through deep learning and the model was evaluated by the verification data. The performance was compared in terms of the accuracy, precision, recall, F1 score through MIT-BIH arrhythmia database. The achieved scores indicate 92.3%, 90.98%, 92.20%, 90.72% in terms of the accuracy, precision, recall, F1 score, respectively.

• The Korean Journal of Applied Statistics
• /
• v.35 no.2
• /
• pp.189-201
• /
• 2022

In this study, we propose factor augmentation to improve forecasting power of cryptocurrency return. We consider financial and economic variables as well as psychological aspect for possible factors. To be more specific, financial and economic factors are obtained by applying principal factor analysis. Psychological factor is summarized by news sentiment analysis. We also visualize such factors through impulse response analysis. In the modeling perspective, we consider ARIMAX as the classical model, and random forest and deep learning to accommodate nonlinear features. As a result, we show that factor augmentation reduces prediction error and the GRU performed the best amongst all models considered.