• Journal of Information Processing Systems
• /
• v.19 no.5
• /
• pp.673-687
• /
• 2023

Convolutional neural network-based deep learning technology is the most commonly used in image identification, but it requires large-scale data for training. Therefore, application in specific fields in which data acquisition is limited, such as in the military, may be challenging. In particular, the identification of ground weapon systems is a very important mission, and high identification accuracy is required. Accordingly, various studies have been conducted to achieve high performance using small-scale data. Among them, the ensemble method, which achieves excellent performance through the prediction average of the pre-trained models, is the most representative method; however, it requires considerable time and effort to find the optimal combination of ensemble models. In addition, there is a performance limitation in the prediction results obtained by using an ensemble method. Furthermore, it is difficult to obtain the ensemble effect using models with imbalanced classification accuracies. In this paper, we propose a transfer learning-based feature fusion technique for heterogeneous models that extracts and fuses features of pre-trained heterogeneous models and finally, fine-tunes hyperparameters of the fully connected layer to improve the classification accuracy. The experimental results of this study indicate that it is possible to overcome the limitations of the existing ensemble methods by improving the classification accuracy through feature fusion between heterogeneous models based on transfer learning.

• Phonetics and Speech Sciences
• /
• v.11 no.4
• /
• pp.89-96
• /
• 2019

In this paper, we propose a deep learning architecture that can effectively detect speech segmentation in broadcast contents. We also propose a multi-scale time-dilated layer for learning the temporal changes of feature vectors. We implement several comparison models to verify the performance of proposed model and calculated the frame-by-frame F-score, precision, and recall. Both the proposed model and the comparison model are trained with the same training data, and we train the model using 32 hours of Korean broadcast data which is composed of various genres (drama, news, documentary, and so on). Our proposed model shows the best performance with F-score 91.7% in Korean broadcast data. The British and Spanish broadcast data also show the highest performance with F-score 87.9% and 92.6%. As a result, our proposed model can contribute to the improvement of performance of speech detection by learning the temporal changes of the feature vectors.

• Journal of the Korean Physical Society
• /
• v.80
• /
• pp.30-36
• /
• 2022

Holographic model of massive scalar field with its self-interaction λϕⁿ in AdS space is able to give a logarithmic scale dependence to marginal multi-trace deformation couplings on its dual conformal field theory, where λ is the self-interaction coupling of the scalar field, ϕ, and n is an integral number. In arXiv:1501.06664, the authors realize this feature by looking at bulk scalar solutions near AdS boundary imposing a specific boundary condition between the coefficients of non-normalizable and normalizable modes of the scalar field excitations. We study the same holographic model to see scale dependence of marginal deformations on the dual conformal field theory by employing completely different method: holographic Wilsonian renormalization group. We solve Hamilton-Jacobi equation derived from the holographic model of massive scalar with λϕⁿ interaction and obtain the solution of marginal multi-trace deformations up to the leading order in λ. It turns out that the solution of marginal multi-trace deformation also presents logarithmic behavior in energy scale near UV region.

• Proceedings of the IEEK Conference
• /
• 2000.09a
• /
• pp.375-378
• /
• 2000

In this paper, we propose a novo] content-based image retrieval system using both Hidden Markov Model(HMM) and an improved chain code. The Gaussian Mixture Model(GMM) is applied to statistically model a color information of the image, and Deterministic Annealing EM(DAEM) algorithm is employed to estimate the parameters of GMM. This result is used to segment the given image. We use an improved chain code, which is invariant to rotation, translation and scale, to extract the feature vectors of the shape for each image in the database. These are stored together in the database with each HMM whose parameters (A, B, $\pi$) are estimated by Baum-Welch algorithm. With respect to feature vector obtained in the same way from the query image, a occurring probability of each image is computed by using the forward algorithm of HMM. We use these probabilities for the image retrieval and present the highest similarity images based on these probabilities.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• v.26 no.1
• /
• pp.133-140
• /
• 2002

This article presents an application of a large-scale structural mixing model(Broadwell et at. 1984) to the blowout of turbulent reacting cross flow jets. Experimental observations, therefore, aim to identify the existence of large-scale vortical structure exerting an important effect upon the flame stabilization. In the analysis of common stability curve, it is seen that the phenomenon of blowout are only related to the mixing time scale of the two flows. The most notable observation is that the blowout distance is traced at a fixed positions according to the velocity ratio at all times. Measurements of the lower blowout limits in the liftable flame are qualitatively in agreement with the blowout parameter $\xi$, proposed by Broadwell et al. Good agrement between the results calculated by a modified blowout parameter $\xi$'and the present experimental results confirms the important effect of large-scale structure in the stabilization feature of blowout.

• Journal of Korean Society on Water Environment
• /
• v.24 no.4
• /
• pp.488-498
• /
• 2008

Soil and Water Assessment Tool (SWAT) model have been widely used in simulating hydrology and water quality analysis at watershed scale. The SWAT model extracts topographic feature using the Digital Elevation Model (DEM) for hydrology and pollutant generation and transportation within watershed. Use of various DEM cell size in the SWAT leads to different results in extracting topographic feature for each subwatershed. So, it is recommended that model users use very detailed spatial resolution DEM for accurate hydrology analysis and water quality simulation. However, use of high resolution DEM is sometimes difficult to obtain and not efficient because of computer processing capacity and model execution time. Thus, the SWAT Topographic Feature Extraction Error (STOPFEE) Fix module, which can extract topographic feature of high resolution DEM from low resolution and updates SWAT topographic feature automatically, was developed and evaluated in this study. The analysis of average slope vs. DEM cell size revealed that average slope of watershed increases with decrease in DEM cell size, finer resolution of DEM. This falsification of topographic feature with low resolution DEM affects soil erosion and sediment behaviors in the watershed. The annual average sediment for Soyanggang-dam watershed with DEM cell size of 20 m was compared with DEM cell size of 100 m. There was 83.8% difference in simulated sediment without STOPFEE module and 4.4% difference with STOPFEE module applied although the same model input data were used in SWAT run. For Imha-dam watershed, there was 43.4% differences without STOPFEE module and 0.3% difference with STOPFEE module. Thus, the STOPFEE topographic database for Soyanggang-dam watershed was applied for Chungju-dam watershed because its topographic features are similar to Soyanggang-dam watershed. Without the STOPFEE module, there was 98.7% difference in simulated sediment for Chungju-dam watershed for DEM cell size of both 20 m and 100 m. However there was 20.7% difference in simulated sediment with STOPFEE topographic database for Soyanggang-dam watershed. The application results of STOPFEE for three watersheds showed that the STOPFEE module developed in this study is an effective tool to extract topographic feature of high resolution DEM from low resolution DEM. With the STOPFEE module, low-capacity computer can be also used for accurate hydrology and sediment modeling for bigger size watershed with the SWAT. It is deemed that the STOPFEE module database needs to be extended for various watersheds in Korea for wide application and accurate SWAT runs with lower resolution DEM.

• Ocean Systems Engineering
• /
• v.12 no.2
• /
• pp.143-157
• /
• 2022

In this study, the nominal wake field of a semi-displacement type high-speed vessel was computed at full scale by using CFD (Computational Fluid Dynamics) and GEOSIM-based approaches. A scale effect investigation on nominal wake field of benchmark Athena vessel was performed with two models which have different model lengths. The members of the model family have the same Fr number but different Re numbers. The spatial components of nominal wake field have been analyzed by considering the axial, radial and tangential velocities for models at different scales. A linear feature has been found for radial and tangential components while a nonlinear change has been obtained for axial velocity. Taylor wake fraction formulation was also computed by using the axial wake velocities and an extrapolation technique was carried out to get the nonlinear fit of nominal wake fraction. This provides not only to observe the change of nominal wake fraction versus scale ratios but also to estimate accurately the wake fraction at full-scale. Extrapolated full-scale nominal wake fractions by GEOSIM-based approach were compared with the full-scale CFD result, and a very good agreement was achieved. It can be noted that the GEOSIM-based extrapolation method can be applied for estimation of the nominal wake fraction of semi-displacement type high-speed vessels.

• International Journal of Ocean Engineering and Technology Speciallssue:Selected Papers
• /
• v.6 no.1
• /
• pp.15-21
• /
• 2003

A common feature in the three-dimensional numerical model experiments of coastal discharge with simplified model and idealized external forcings is investigated. The velocity fields due to the buoyancy and flaw flux, are spreaded radiately and the surface velocites are much greater than the homegeneous discharges. The coastal dischargd due to the Coriolis force and flaw flux are shaped a anticyclical gyre (clockwise) and determined the scale of the gyre in the coastal zone, respectively. The bottom topography restricts a outward extention of the coastal fronts and it accelerates a southward flow.

• Journal of the Korea Society of Computer and Information
• /
• v.13 no.4
• /
• pp.51-58
• /
• 2008

In this paper, we propose a place modeling based on the distribution of scale-invariant features, and a place recognition method that recognizes places by comparing the place model in a database with the extracted features from input data. The proposed method is based on the assumption that every place can be represented by unique feature distributions that are distinguishable from others. The proposed method uses global information of each place where one place is represented by one distribution model. Therefore, the main contribution of the proposed method is that the time cost corresponding to the increase of the number of places grows linearly without increasing exponentially. For the performance evaluation of the proposed method, the different number of frames and the different number of features are used, respectively. Empirical results illustrate that our approach achieves better performance in space and time cost comparing to other approaches. We expect that the Proposed method is applicable to many ubiquitous systems such as robot navigation, vision system for blind people, wearable computing, and so on.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.2
• /
• pp.33-42
• /
• 2022

In this paper, we proposed a deep learning model based on multi-scale dilated convolution feature fusion for the segmentation of black ice region in road image to send black ice warning to drivers in real time. In the proposed multi-scale dilated convolution feature fusion network, different dilated ratio convolutions are connected in parallel in the encoder blocks, and different dilated ratios are used in different resolution feature maps, and multi-layer feature information are fused together. The multi-scale dilated convolution feature fusion improves the performance by diversifying and expending the receptive field of the network and by preserving detailed space information and enhancing the effectiveness of diated convolutions. The performance of the proposed network model was gradually improved with the increase of the number of dilated convolution branch. The mIoU value of the proposed method is 96.46%, which was higher than the existing networks such as U-Net, FCN, PSPNet, ENet, LinkNet. The parameter was 1,858K, which was 6 times smaller than the existing LinkNet model. From the experimental results of Jetson Nano, the FPS of the proposed method was 3.63, which can realize segmentation of black ice field in real time.