• ETRI Journal
• /
• v.45 no.5
• /
• pp.822-835
• /
• 2023

Recent semantic segmentation frameworks usually combine low-level and high-level context information to achieve improved performance. In addition, postlevel context information is also considered. In this study, we present a Context ReFinement Network (CRFNet) and its training method to improve the semantic predictions of segmentation models of the encoder-decoder structure. Our study is based on postprocessing, which directly considers the relationship between spatially neighboring pixels of a label map, such as Markov and conditional random fields. CRFNet comprises two modules: a refiner and a combiner that, respectively, refine the context information from the output features of the conventional semantic segmentation network model and combine the refined features with the intermediate features from the decoding process of the segmentation model to produce the final output. To train CRFNet to refine the semantic predictions more accurately, we proposed a sequential training scheme. Using various backbone networks (ENet, ERFNet, and HyperSeg), we extensively evaluated our model on three large-scale, real-world datasets to demonstrate the effectiveness of our approach.

• Journal of Information Technology Services
• /
• v.22 no.5
• /
• pp.87-98
• /
• 2023

Traditionally, tooth restoration has been carried out by replicating teeth using plaster-based materials. However, recent technological advances have simplified the production process through the introduction of computer-aided design(CAD) systems. Nevertheless, dental restoration varies among individuals, and the skill level of dental technicians significantly influences the accuracy of the manufacturing process. To address this challenge, this paper proposes an approach to designing personalized tooth restorations using Generative Adversarial Network(GAN), a widely adopted technique in computer vision. The primary objective of this model is to create customized dental prosthesis for each patient by utilizing 3D data of the specific teeth to be treated and their corresponding opposite tooth. To achieve this, the 3D dental data is converted into a depth map format and used as input data for the GAN model. The proposed model leverages the network architecture of Pixel2Style2Pixel, which has demonstrated superior performance compared to existing models for image conversion and dental prosthesis generation. Furthermore, this approach holds promising potential for future advancements in dental and implant production.

• Korean Journal of Remote Sensing
• /
• v.38 no.5_3
• /
• pp.939-951
• /
• 2022

The most important factor in designing autonomous driving systems is to recognize the exact location of the vehicle within the surrounding environment. To date, various sensors and navigation systems have been used for autonomous driving systems; however, all have limitations. Therefore, the need for high-definition (HD) maps that provide high-precision infrastructure information for safe and convenient autonomous driving is increasing. HD maps are drawn using three-dimensional point cloud data acquired through a mobile mapping system (MMS). However, this process requires manual work due to the large numbers of points and drawing layers, increasing the cost and effort associated with HD mapping. The objective of this study was to improve the efficiency of HD mapping by segmenting semantic information in an MMS point cloud into six classes: roads, curbs, sidewalks, medians, lanes, and other elements. Segmentation was performed using various machine learning techniques including random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN), and gradient-boosting machine (GBM), and 11 variables including geometry, color, intensity, and other road design features. MMS point cloud data for a 130-m section of a five-lane road near Minam Station in Busan, were used to evaluate the segmentation models; the average F1 scores of the models were 95.43% for RF, 92.1% for SVM, 91.05% for GBM, and 82.63% for KNN. The RF model showed the best segmentation performance, with F1 scores of 99.3%, 95.5%, 94.5%, 93.5%, and 90.1% for roads, sidewalks, curbs, medians, and lanes, respectively. The variable importance results of the RF model showed high mean decrease accuracy and mean decrease gini for XY dist. and Z dist. variables related to road design, respectively. Thus, variables related to road design contributed significantly to the segmentation of semantic information. The results of this study demonstrate the applicability of segmentation of MMS point cloud data based on machine learning, and will help to reduce the cost and effort associated with HD mapping.

• Journal of dental hygiene science
• /
• v.14 no.2
• /
• pp.94-100
• /
• 2014

The purpose of this study was to evaluate the accuracy of digitized elastomeric impression materials of crown abutment, using non-contact white light scanner and virtual three-dimensional superimpositional analysis. The stone models and impressions were digitized white light scanner to create three-dimensional surface models. Stone models were used as CAD reference model (CRM). The resulting point clouds (ASC file) from digitization of impressions using converting software. Discrepancies between the points in the point clouds and CRM were measured by superimpositional software. Mean and standard deviation of values of discrepancies were analyzed by one-way ANOVA and multiple comparison (${\alpha}=0.05$). The mean discrepancy between the impressions for the extra-light body (XLB), light body (LB), and heavy body (HB) group were $5.10{\pm}1.45{\mu}m$, $6.30{\pm}1.87{\mu}m$, $9.80{\pm}1.52{\mu}m$, respectively. The different impression materials affected the digitization of impressions significantly (p<0.05). As a result, digitization of elastomeric impression materials on dental scanner was influenced by material sort.

• Spatial Information Research
• /
• v.18 no.4
• /
• pp.1-11
• /
• 2010

Unlike the previous studies which have focused on 2-dimensional urban characteristics, this paper presents statistical models explaining urban heat island(UHI) effect by 3-dimensional urban morphologic information and addresses its policy implications. 3~dimensional informations of Columbus, Ohio arc captured from LiDAR data and building boundary informations are extracted from a building digital map, Finally NDV[ and temperature data are calculated by manipulating band 3, band 4, and thermal hand of LandSat images. Through complicated data processing, 6 independent variables(building surface area, building volume, height to width ratio, porosity, plan surface area) are introduced in simple and multiple linear regression models. The regression models are specified by Box-Tidwell method, finding the power to which the independent variable needs to raised to be in a linearity. Porosity, NDVI, and building surface area are carefully chosen as explanatory variables in the final multiple regression model, which explaining about 57% of the variability in temperatures. On reducing UHI, various implications of the results give guidelines to policy-making in open space, roof garden, and vertical garden management.

• Journal of Environmental Impact Assessment
• /
• v.22 no.3
• /
• pp.219-230
• /
• 2013

Species distribution modeling is one of the most effective habitat analysis methods for wildlife conservation. This study was for evaluating the suitability of species distribution to distance between forest patches in Seoul city using tits. We analyzed the distribution of the four species of tits: varied tit (Parus varius), marsh tit (P. palustris), great tit (P. major) and coal tit (P. ater), using the landscape indexes and connectivity indexes, and compared the resulting suitability indexes from 100m to 1,000m. As factors affecting to the distribution of tits, we calculated landscape indices by separating them into intra-patch indices (i.e. logged patch area (PA), area-weighted mean patch shape index (PSI), tree rate (TR)) and inter-patch indices (i.e. patch degree (PD), patch betweenness (PB), difference probability of connectivity (DPC)), to analyze the internal properties of the patches and their connectivity by tits occurrence data using logistic regression modeling. The models were evaluated by AICc (Akaike Information Criteria with a correction for finite sample sizes) and AUC (Area Under Curve of ROC). The results of AICc and AUC showed DPC, PA, PSI, and TR were important factors of the habitat models for great tit and marsh tit at the level of distance 500~800m. In contrast, habitat models for coal tit and varied tit, which are known as forest interior species, reflected PA, PSI, and TR as intra-patch indices rather than connectivity. These mean that coal tit and varied tit are more likely to find a large circular forest patch than a small and long-shaped forest patch, which are higher rate of forest. Therefore, different strategies are required in order to enhance the habitats of the forest birds, tits, in a region that has fragmented forest patches such as Seoul city. It is important to manage forest interior areas for coal tit and varied tit, which are known as forest interior species and to manage not only forest interior areas but also connectivity of the forest patches in the threshold distance for great tit and marsh tit as adapted species to the urban ecosystem for sustainable ecosystem management.

• Journal of Navigation and Port Research
• /
• v.40 no.1
• /
• pp.27-34
• /
• 2016

For the prevention of marine casualties, international bodies have mainly focused on strengthening ship's stability and design, maritime education and training, and improving maritime traffic environment. Statistics analysis on marine casualties showed that most of casualties occurred in coastal waters, especially by human elements. In order to review the conformity of existing prevention measures with the result of the statistics analysis, the IMO's SHELL model was applied to the established measures. As a result, ergonomic approaches were needed for the prevention of human errors in coastal waters, so that the priority should be given to the interface between ship's operator and navigational environment. For this study, Rasmussen's SRK pyramid, which showed decision making mechanism of human, and the US Coast Guard's investigation manual on marine casualties concerning the collapse of safe maritime transportation system were reviewed, and the merits and demerits within the risk assessment tools such as IWRAP, PAWSA, ES model, PARK model, and NURI model were also studied. Although the effectiveness of the existing risk assessment models was proved in ports and approaching channels, it is concluded that the need of new models for converting Korean seafarers' qualitative risk to quantitative risk was proposed so as to print hazard maps which make seafarers instinctively recognize comparative hazard levels of coastal waters.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.8
• /
• pp.49-59
• /
• 2022

In this paper, we designed a new enzyme function prediction model PSCREM based on a study that compared and evaluated CNN and LSTM/GRU models, which are the most widely used deep learning models in the field of predicting functions and structures using protein sequences in 2020, under the same conditions. Sequence evolution information was used to preserve detailed patterns which would miss in CNN convolution, and the relationship information between amino acids with functional significance was extracted through overlapping RNNs. It was referenced to feature map production. The RNN family of algorithms used in small CNN-RNN models are LSTM algorithms and GRU algorithms, which are usually stacked two to three times over 100 units, but in this paper, small RNNs consisting of 10 and 20 units are overlapped. The model used the PSSM profile, which is transformed from protein sequence data. The experiment proved 86.4% the performance for the problem of predicting the main classes of enzyme number, and it was confirmed that the performance was 84.4% accurate up to the sub-sub classes of enzyme number. Thus, PSCREM better identifies unique patterns related to protein function through overlapped RNN, and Overlapped RNN is proposed as a novel methodology for protein function and structure prediction extraction.

• Journal of the Korean Society of Hazard Mitigation
• /
• v.9 no.4
• /
• pp.71-80
• /
• 2009

Although desire for living without hazardous damages grows these days, threats from natural disasters which we are currently exposed to are quiet different from what we have experienced. To cope with this changing situation, it is necessary to assess the characteristics of the natural disasters. Therefore, the main purpose of this research is to suggest a methodology to estimate the potential property loss and assess the flood risk using a regional regression analysis. Since the flood damage mainly consists of loss of lives and property damages, it is reasonable to express the results of a flood risk assessment with the loss of lives and the property damages that are vulnerable to flood. The regional regression analysis has been commonly used to find relationships between regional characteristics of a watershed and parameters of rainfall-runoff models or probability distribution models. In our research, however, this model is applied to estimate the potential flood damage as follows; 1) a nonlinear model between the flood damage and the hourly rainfall is found in gauged regions which have sufficient damage and rainfall data, and 2) a regression model is developed from the relationship between the coefficients of the nonlinear models and socio-economic indicators in the gauged regions. This method enables us to quantitatively analyze the impact of the regional indicators on the flood damage and to estimate the damage through the application of the regional regression model to ungauged regions which do not have sufficient data. Moreover the flood risk map is developed by Flood Vulnerability Index (FVI) which is equal to the ratio of the estimated flood damage to the total regional property. Comparing the results of this research with Potential Flood Damage (PFD) reported in the Long-term Korea National Water Resources Plan, the exports' mistaken opinions could affect the weighting procedure of PFD, but the proposed approach based on the regional regression would overcome the drawback of PFD. It was found that FVI is highly correlated with the past damage, while PFD does not reflect the regional vulnerabilities.

• Journal of the Korea Computer Graphics Society
• /
• v.6 no.4
• /
• pp.1-7
• /
• 2000

Traditionally, three dimension model s have been used for building virtual worlds, and a data structure called the "scene graph" is often employed to organize these 3D objects in the virtual space. On the other hand, image-based rendering has recently been suggested as a probable alternative VR platform for its photo-realism, however, due to limited interactivity. it has only been used for simple navigation systems. To combine the merits of these two approaches to object/scene representations, this paper proposes for a scene graph structure in which both 3D models and various image-based scenes/objects can be defined. traversed, and rendered together. In fact, as suggested by Shade et al. [1]. these different representations can be used as different LOD's for a given object. For in stance, an object might be rendered using a 3D model at close range, a billboard at an intermediate range. and as part of an environment map at far range. The ultimate objective of this mixed platform is to breath more interactivity into the image based rendered VE's by employing 3D models as well. There are several technical challenges in devising such a platform : designing scene graph nodes for various types of image based techniques, establishing criteria for LOD/representation selection. handling their transition s. implementing appropriate interaction schemes. and correctly rendering the overall scene. Currently, we have extended the scene graph structure of the Sense8's WorldToolKit. to accommodate new node types for environment maps. billboards, moving textures and sprites, "Tour-into-the-Picture" structure, and view interpolated objects. As for choosing the right LOD level, the usual viewing distance and image space criteria are used, however, the switching between the image and 3D model occurs at a distance from the user where the user starts to perceive the object's internal depth. Also. during interaction, regardless of the viewing distance. a 3D representation would be used, if it exists. Finally. we carried out experiments to verify the theoretical derivation of the switching rule and obtained positive results.