• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.6
• /
• pp.629-635
• /
• 2019

In Deep Learning models derivative is implemented by error back-propagation which enables the model to learn the error and update parameters. It can find the global (or local) optimal points of parameters even in the complex models taking advantage of a huge improvement in computing power. However, deliberately generated data points can 'fool' models and degrade the performance such as prediction accuracy. Not only these adversarial examples reduce the performance but also these examples are not easily detectable with human's eyes. In this work, we propose the method to detect adversarial datasets with random noise addition. We exploit the fact that when random noise is added, prediction accuracy of non-adversarial dataset remains almost unchanged, but that of adversarial dataset changes. We set attack methods (FGSM, Saliency Map) and noise level (0-19 with max pixel value 255) as independent variables and difference of prediction accuracy when noise was added as dependent variable in a simulation experiment. We have succeeded in extracting the threshold that separates non-adversarial and adversarial dataset. We detected the adversarial dataset using this threshold.

• The korean journal of orthodontics
• /
• v.35 no.5 s.112
• /
• pp.388-397
• /
• 2005

The purpose of this study was to redefine the cephalometric landmarks in three-dimensional (3D) images, which are used in orthodontic cephalometric radiography, and to evaluate the reproducibility of each landmark for 3D cephalometric analysis. Eighteen CT scans were taken at the Department of Diagnostic Radiology at Seoul National University Dental Hospital and manipulated with V works 4.0(Cybermed Inc., Seoul, Korea). The coordinate system was established using 7 reference points, with no more than 4 points on the same plane. These 7 points were generated as a volume model, the voxel size of which was 4 by 4 by 2 (threshold value=639). The cephalometric landmarks were selected at the multiplanar reformation (MPR) window on the volume mode of V works 4.0. The selected landmarks were exported to V surgery (Cybermed Inc., Seoul, Korea) for the calculation of coordinate values. All the data were taken twice with a lapse of 2 weeks by one investigator The reproducibility of each landmark was $0.17\~1.21mm$ in the x axis, $0.30\~1.53mm$. In the y axis, and $0.27\~1.81mm$ in the z axis. In all three axes, the range of error was similar. These error ranges were acceptable with regards to the pixel space and slice thickness. The most reproducible points were 1 points which were selected on the basis of the volume model. The least reproducible points were J points that were defined by sutures.

• Korean Journal of Remote Sensing
• /
• v.35 no.6_1
• /
• pp.907-917
• /
• 2019

Recently, the demand for high-resolution satellite images increases in many fields such as land monitoring and terrain analysis. Therefore, the need for geometric correction is increasing. As an automatic precision geometric correction method, there is a method of automatically extracting the GCP by matching between the GCP Chip and the satellite image. For automatic precision geometric correction, the success rate of matching GCP Chip and satellite image is important. Therefore, it is important to evaluate the matching performance of the manufactured GCP Chip. In order to evaluate the matching performance of GCP Chips, a total of 3,812 GCP Chips in South Korea were used as experimental data. The GCP Chip matching results of KOMPSAT-3A and Google Map showed similar matching results. Therefore, we determined that Google Map satellite imagery could replace high-resolution satellite imagery. Also, presented a method using center point and error radius of Google Map to reduce the time required to verify matching performance. As a result, it is best to set the optimum error radius to 8.5m. Evaluated the matching performance of GCP Chips in South Korea using Google Maps. And verified matching result using presented method. As a result, the GCP Chip s in South Korea had a matching success rate of about 94%. Also, the main matching failure factors were analyzed by matching failure GCP Chips. As a result, Except for GCP Chips that need to be remanufactured, the remaining GCP Chips can be used for the automatic geometric correction of satellite images.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.28 no.2
• /
• pp.247-254
• /
• 2010

In this paper, a new approach has been studied correcting the geometric distortion of SPOT 4 imagery. Two new equations were induced by the relationship between satellite and the Earth in the space. line-of-sight (LOS) vector adjustment model for SPOT 4 imagery was implemented in this study. This model is to adjust LOS vector under the assumption that the orbital information of satellite provided by receiving station is uncertain and this uncertainty makes a constant error over the image. This model is verified using SPOT 4 satellite image with high look angle and thirty five ground points, which include 10 GCPs(Ground Control Points) and 25 check points, measured by the GPS. In total thirty five points, the geometry of satellite image calculated by given satellite information(such as satellite position, velocity, attitude and look angles, etc) from SPOT 4 satellite image was distorted with a constant error. Through out the study, it was confirmed that the LOS vector adjustment model was able to be applied to SPOT4 satellite image. Using this model, RMSEs (Root Mean Square Errors) of twenty five check points taken by increasing the number of GCPs from two to ten were less than one pixel. As a result, LOS vector adjustment model could efficiently correct the geometry of SPOT4 images with only two GCPs. This method also is expected to get good results for the different satellite images that are similar to the geometry of SPOT images.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.45 no.5
• /
• pp.1-8
• /
• 2008

The belief propagation method that has been studied recently yields good performance in disparity extraction. The method in which a target function is modeled as an energy function based on Markov random field(MRF), solves the stereo matching problem by finding the disparity to minimize the energy function. MRF models provide robust and unified framework for vision problem such as stereo and image restoration. the belief propagation method produces quite correct results, but it has difficulty in real time implementation because of higher computational complexity than other stereo methods. To relieve this problem, in this paper, we propose a fast algorithm of the belief propagation method. Energy function consists of a data term and a smoothness tern. The data term usually corresponds to the difference in brightness between correspondences, and smoothness term indicates the continuity of adjacent pixels. Smoothness information is created from messages, which are assigned using four different message arrays for the pixel positions adjacent in four directions. The processing time for four message arrays dominates 80 percent of the whole program execution time. In the proposed method, we propose an algorithm that dramatically reduces the processing time require in message calculation, since the message.; are not produced in four arrays but in a single array. Tn the last step of disparity extraction process, the messages are called in the single integrated array and this algorithm requires 1/4 computational complexity of the conventional method. Our method is evaluated by comparing the disparity error rates of our method and the conventional method. Experimental results show that the proposed method remarkably reduces the execution time while it rarely increases disparity error.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.16 no.6
• /
• pp.90-100
• /
• 2017

The purpose of this study is to detect and label the vehicles using the drone images as a way to overcome the limitation of the existing point and section detection system and vehicle gap estimation on Arterial road. In order to select the appropriate time zone, position, and altitude for the acquisition of the drone image data, the final image data was acquired by shooting under various conditions. The vehicle was detected by applying mixed Gaussian, image binarization and morphology among various image analysis techniques, and the vehicle was labeled by applying Kalman filter. As a result of the labeling rate analysis, it was confirmed that the vehicle labeling rate is 65% by detecting 185 out of 285 vehicles. The gap was calculated by pixel unitization, and the results were verified through comparison and analysis with Daum maps. As a result, the gap error was less than 5m and the mean error was 1.67m with the preceding vehicle and 1.1m with the following vehicle. The gaps estimated in this study can be used as the density of the urban roads and the criteria for judging the service level.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.36 no.4
• /
• pp.295-303
• /
• 2018

The stereo camera system has a fixed baseline and therefore has a constant scale. However, it is difficult to measure the actual three-dimensional coordinate since the scale is not fixed when relative orientation parameters are determined through the key-point matching in the stereo image each time. Therefore, the purpose of this study was to perform the stereo camera calibration that simultaneously determines the internal characteristics of the left and right cameras and the camera relationship between them using the modified collinearity equation and compared it with the two independent single cameras calibration. In the experiment using the images taken at close range, the RMSE (Root Mean Square Error) of ${\pm}0.014m$ was occurred when the three dimensional distances were compared in the single calibration results. On the other hand, the accuracy of the three-dimensional distance of the stereo camera calibration was better because the stereo camera results were almost no error compared to the results from two single cameras. In the comparison of the epipolar images, the RMSE of the stereo camera was 0.3 pixel more than that of the two single cameras, but the effect was not significant.

• Korean Journal of Remote Sensing
• /
• v.23 no.6
• /
• pp.537-546
• /
• 2007

In hyperspectral remote sensing, linear spectral mixture model is a common procedure decomposing into the components of a mixed pixel and estimating the fraction of each end-member. Although linear spectral mixture model is frequently used in geology and mineral mapping because this model is simple and easy to apply, this model is not always valid in forest and urban area having rather complex structure. This study aims to analyze possible error for applying linear spectral mixture model. For the study, we measured laboratory spectra of mixture sample, having various materials, fractions, distributions. The accuracy of linear mixture model is low with the mixture sample having similar fraction because the multi-scattering between components is maximum. Additionally, this multi-scattering is related to the types, fraction, and distribution of components. Further analysis is necessary to quantify errors from linear spectral mixture model.

• Korean Journal of Remote Sensing
• /
• v.37 no.5_1
• /
• pp.1013-1027
• /
• 2021

Images taken with KOMPSAT-3 have additional NIR and PAN bands, as well as RGB regions of the visible ray band, compared to imagestaken with a standard camera. Furthermore, electrical and optical properties must be considered because a wide radius area of approximately 17 km or more is photographed at an altitude of 685 km above the ground. In other words, the camera sensor of KOMPSAT-3 is distorted by each CCD pixel, characteristics of each band,sensitivity and time-dependent change, CCD geometry. In order to solve the distortion, correction of the sensors is essential. In this paper, we propose a method for detecting uniform regions in side-slider-based KOMPSAT-3 images using segment-based noise analysis. After detecting a uniform area with the corresponding algorithm, a correction table was created for each sensor to apply the non-uniformity correction algorithm, and satellite image correction was performed using the created correction table. As a result, the proposed method reduced the distortion of the satellite image,such as vertical noise, compared to the conventional method. The relative radiation accuracy index, which is an index based on mean square error (RA) and an index based on absolute error (RE), wasfound to have a comparative advantage of 0.3 percent and 0.15 percent, respectively, over the conventional method.

• The Korean Journal of Nuclear Medicine
• /
• v.32 no.5
• /
• pp.414-424
• /
• 1998

Purpose: Cross-modality coregistration of positron emission tomography (PET) and magnetic resonance imaging (MR) could enhance the clinical information. In this study we propose a refined technique to improve the robustness of registration, and to implement more realistic visualization of the coregistered images. Materials and Methods: Using the sinogram of PET emission scan, we extracted the robust head boundary and used boundary-enhanced PET to coregister PET with MR. The pixels having 10% of maximum pixel value were considered as the boundary of sinogram. Boundary pixel values were exchanged with maximum value of sinogram. One hundred eighty boundary points were extracted at intervals of about 2 degree using simple threshold method from each slice of MR images. Best affined transformation between the two point sets was performed using least square fitting which should minimize the sum of Euclidean distance between the point sets. We reduced calculation time using pre-defined distance map. Finally we developed an automatic coregistration program using this boundary detection and surface matching technique. We designed a new weighted normalization technique to display the coregistered PET and MR images simultaneously. Results: Using our newly developed method, robust extraction of head boundary was possible and spatial registration was successfully performed. Mean displacement error was less than 2.0 mm. In visualization of coregistered images using weighted normalization method, structures shown in MR image could be realistically represented. Conclusion: Our refined technique could practically enhance the performance of automated three dimensional coregistration.