• 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.

• Smart Media Journal
• /
• v.2 no.4
• /
• pp.60-65
• /
• 2013

In this paper, we propose a high-resolution disparity map generation method using a low-resolution time-of-flight (TOF) depth camera and color camera. The TOF depth camera is efficient since it measures the range information of objects using the infra-red (IR) signal in real-time. It also quantizes the range information and provides the depth image. However, there are some problems of the TOF depth camera, such as noise and lens distortion. Moreover, the output resolution of the TOF depth camera is too small for 3D applications. Therefore, it is essential to not only reduce the noise and distortion but also enlarge the output resolution of the TOF depth image. Our proposed method generates a depth map for a color image using the TOF camera and the color camera simultaneously. We warp the depth value at each pixel to the color image position. The color image is segmented using the mean-shift segmentation method. We define a cost function that consists of color values and segmented color values. We apply a weighted average filter whose weighting factor is defined by the random walk probability using the defined cost function of the block. Experimental results show that the proposed method generates the depth map efficiently and we can reconstruct good virtual view images.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.28 no.1C
• /
• pp.1-8
• /
• 2003

In this paper, performance of the CPC(complex phase code) which is recently proposed as a practical phase encoding method for phase-code multiplexes holographic memory system is comparatively analyzed with those of the conventional phase codes such as PR(pure random code), RCE(random code with equality), WHM(Walsh Hadamard Matrix). In computer simulation, the size of an address bean is fixed at 32$\times$32 pixels and 0%-25% phase-error ratio in a pixel are intentionally added to the real phase values to consider the nonlinear phase-modulation characteristics of the practical spatial light modulator. From comparative analysis of crosstalks and signal-to-noise ratios for these phase codes by calculating auto-correlation and cross-correlation, it is found that the CPC have the lowest cross-correlation mean value of 0.021, the lowest standard deviation of 0.0113 and the highest signal-to-noise ratio(SNR) of 27.4 among the four types of phase code. In addition, from the calculation of the number of all possible address beams for these four types of phase code as the size of the address beam is fixed to 3232 pixels, the CPC is found to have 6.334$\times$10$^{49}$ address beams, which are relatively higher number than that of the conventional phase codes.

• Korean Journal of Digital Imaging in Medicine
• /
• v.14 no.2
• /
• pp.1-8
• /
• 2012

Purpose : We aim at presenting the optimum radiologic factor through the evaluation of dose variation and of image quality through the use of a grid in Humerus examination and the change of dose because of the change of radiologic factor. Materials and Methods : We divided it in 3 cases: when using a grid or not and when using IP(Image Plate) in a digital system. Also, as fixing kVp to 70kVp it changed mAs, and fixing mAs to 10 it changed kVp, we put up resolution chart and Burger rose phantom on the acrylic phantom of 7cm (the same level of Humerus) to evaluate the dose and image. We used Image J program to evaluate the quantitative resolution of the obtained image, and made the qualitative evaluation and statistical analysis of the image saved in PACS for 20 radiologic technologist with more than 10 years of experience in order of evaluate its contrast. We used SPSS10(SPSS Inc. Chicago, Illinois) for statistical analysis. Results : We observed the analytic result of resolution by the change of kVp that it was $4.539dGycm^2$ in 60kVp and $757.472dGycm^2$ in 75kVp, which increased about 64.6% of dose, while for the resolution it had the pixel value 30.7% better with 851 in 60kVp than 651 in 75kVp. Also, we analyzed the result of resolution by the change of mAs that it was $3.106dGycm^2$ in 5mAs, and $12.470dGycm^2$ in 20mAs, which increased about 400% of dose, while for the resolution DR had 678 in 5mAs, and 724 in 20mAs that increased about 6.8% of resolution. We made the qualitative evaluation of contrast by the change of kVp that DR showed the higher quality than CR, but the contrast by the change of kVp had no special different at the moment of visual evaluation, nor statistically significant difference(P>0.05). We observed the qualitative evaluation of contraste by the change of mAs that the contrast increased as DR increased mAs, and had statistically significant difference(P<0.05). On the other hand, CR had no significant difference for more than 10mAs nor statistically significant difference(P>0.05). Conclusion : In case of some patients with radiographic exposure by the repeated examination such as emergent patient or Follow up patient, they are considered to try to limit the use of a grid, to set kVp under 65kVp in fixed mode, to select less than 10mAs and to reduce the possibility of patient being bombed.

• Spatial Information Research
• /
• v.22 no.3
• /
• pp.47-58
• /
• 2014

The credible measurement, reporting and verification (MRV) is among the most critical elements in UN-REDD (United Nations programme on Reducing Emissions from Deforestation and forest Degradation in Developing Countries). This study is intended to explore MRV potential in terms of UN-REDD opportunity cost estimation using satellite image for Mt. Geum-gang of North Korea. A visual interpretation were conducted to evaluate MRV conditions by sub-dividing or decomposing the images with different pixel size into a three types of hierarchical tree structure that helps dealing with spatial variability within each subarea. The permanent record of standard satellite remote sensing system demonstrated its capability of presenting area-wide visual evidences of MRV conditions in Mt. Geum-gang (such as the identification of forested area, degradation trends for forest space, three types of hierarchical land-cover and land use tree structure, carbon density in the landscape). Satellite data could be accepted as legally binding proof when it comes to REDD opportunity cost estimation since several cases exist where remote sensing has been used as legal evidence in ICJ (International Court of Justice) and UN resolution. It doesn't seem very difficult to comply with MRV requirements for UN-REDD opportunity cost calculation due to the probative value of satellite data. It is anticipated that this research output could be used as a valuable reference for Korea-based enterprises exploring REDD project sites and the carbon traders to ensure MRV potentials using satellite image in UN-REDD Opportunity Cost estimation.

• KIPS Transactions on Software and Data Engineering
• /
• v.1 no.2
• /
• pp.109-114
• /
• 2012

Honeycombs are dense structures that small cysts, which generally have about 2~10 mm in diameter, are surrounded by the wall of fibrosis. When honeycomb is found in the patients, the incidence of acute exacerbation is generally very high. Thus, the observation and quantitative measurement of honeycomb are considered as a significant marker for clinical diagnosis. In this point of view, we propose an automatic segmentation method using morphological image processing and assessment of the degree of clustering techniques. Firstly, image noises were removed by the Gaussian filtering and then a morphological dilation method was applied to segment lung regions. Secondly, honeycomb cyst candidates were detected through the 8-neighborhood pixel exploration, and then non-cyst regions were removed using the region growing method and wall pattern testing. Lastly, final honeycomb regions were segmented through the extraction of dense regions which are consisted of two or more cysts using cluster analysis. The proposed method applied to 80 High resolution computed tomography (HRCT) images and achieved a sensitivity of 89.4% and PPV (Positive Predictive Value) of 72.2%.

• Korean Journal of Optics and Photonics
• /
• v.33 no.4
• /
• pp.146-158
• /
• 2022

The detector of a focal-plane-array mid-wave infrared (MWIR) camera has different response characteristics for each detector pixel, resulting in nonuniformity between detector pixels. In addition, image nonuniformity occurs due to heat generation inside the camera during operation. To solve this problem, in the process of camera manufacturing it is common to use a gain-and-offset table generated from a blackbody to correct the difference between detector pixels. One method of correcting nonuniformity due to internal heat generation during the operation of the camera generates a new offset value based on input frame images. This paper proposes a technique for dividing an input image into block image patches and generating offset values using only homogeneous patches, to correct the nonuniformity that occurs during camera operation. The proposed technique may not only generate a nonuniformity-correction offset that can prevent motion marks due to camera-gaze movement of the acquired image, but may also improve nonuniformity-correction performance with a small number of input images. Experimental results show that distortion such as flow marks does not occur, and good correction performance can be confirmed even with half the number of input images or fewer, compared to the traditional method.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1301-1314
• /
• 2022

Normalized Difference Vegetation Index (NDVI) is utilized as an indicator to represent the vegetation condition on the land surface in various applications such as land cover, crop yield, agricultural drought, soil moisture, and forest disaster. However, satellite optical sensors for visible and infrared rays cannot see through the clouds, so the NDVI of the cloud pixel is not a valid value for the land surface. This study proposed a real-time correction of the underestimation noise for GEO-KOMPSAT-2A (GK2A) daily NDVI and made sure its feasibility through the quantitative comparisons with Moderate Resolution Imaging Spectroradiometer (MODIS) NDVI and the qualitative interpretation of time-series changes. The underestimation noise was effectively corrected by the procedures such as the time-series correction considering vegetation phenology, the outlier removal using long-term climatology, and the gap filling using rigorous statistical methods. The correlation with MODIS NDVI was higher, and the difference was lower, showing a 32.7% improvement compared to the original NDVI product. The proposed method has an extensibility for use in other satellite products with some modification.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.9
• /
• pp.1-12
• /
• 2022

Traditional methods of measuring corneal ulcers were difficult to present objective basis for diagnosis because of the subjective judgment of the medical staff through photographs taken with special equipment. In this paper, we propose a method to detect the ulcer area on a pixel basis in corneal ulcer images using a semantic segmentation model. In order to solve this problem, we performed the experiment to detect the ulcer area based on the DeepLab model which has the highest performance in semantic segmentation model. For the experiment, the training and test data were selected and the backbone network of DeepLab model which set as Xception and ResNet, respectively were evaluated and compared the performances. We used Dice similarity coefficient and IoU value as an indicator to evaluate the performances. Experimental results show that when 'crop & resized' images are added to the dataset, it segment the ulcer area with an average accuracy about 93% of Dice similarity coefficient on the DeepLab model with ResNet101 as the backbone network. This study shows that the semantic segmentation model used for object detection also has an ability to make significant results when classifying objects with irregular shapes such as corneal ulcers. Ultimately, we will perform the extension of datasets and experiment with adaptive learning methods through future studies so that they can be implemented in real medical diagnosis environment.

• Proceedings of the Korean Society of Medical Physics Conference
• /
• 2003.09a
• /
• pp.67-67
• /
• 2003

It is crucial to minimize setup errors of a cancer treatment machine using a high energy and to perform precise radiation therapy. Usually, port film has been used for verifying errors. The Korea Cancer Center Hospital (KCCH) has manufactured digital electronic portal imaging device (EPID) system to verify treatment machine errors as a Quality Assurance (Q.A) tool. This EPID was consisted of a metal/fluorescent screen, 45$^{\circ}$ mirror, a camera and an image grabber and could display the portal image with near real time KIRAMS has also made the acrylic phantom that has lead line of 1mm width for ligh/radiation field congruence verification and reference points phantom for using as an isocenter on portal image. We acquired portal images of 10$\times$10cm field size with this phantom by EPID and portal film rotating treatment head by 0.3$^{\circ}$, 0.6$^{\circ}$ and 0.9$^{\circ}$. To check field size, we acquired portal images with 18$\times$18cm, 19$\times$19cm and 20$\times$20cm field size with collimator angle 0$^{\circ}$ and 0.5$^{\circ}$ individually. We have performed Flatness comparison by displaying the line intensity from EPID and film images. The 0.6$^{\circ}$ shift of collimator angle was easily observed by edge detection of irradiated field size on EPID image. To the extent of one pixel (0.76mm) difference could be detected. We also have measured field size by finding optimal threshold value, finding isocenter, finding field edge and gauging distance between isocenter and edge. This EPID system could be used as a Q.A tool for checking field size, light/radiation congruence and flatness with a developed video based EPID.