• Annual Conference of KIPS
• /
• 2022.11a
• /
• pp.520-522
• /
• 2022

의료 인공지능은 특정 진단에서 높은 정확도를 보이지만 모델의 신뢰성 문제로 인해 활발하게 쓰이지 못하고 있다. 이에 따라 인공지능 모델의 진단에 대한 원인 설명의 필요성이 대두되었고 설명가능한 의료 인공지능에 관한 연구가 활발히 진행되고 있다. 하지만 MRI 등 의료 영상 인공지능 분야에서 주로 진행되고 있으며, 이미지 형태가 아닌 전자의무기록 데이터 (Electronic Health Record, EHR) 를 기반으로 한 모델의 설명가능성 연구는 EHR 데이터 자체의 복잡성 때문에 활발하게 진행 되지 않고 있다. 본 논문에서는 전자의무기록 데이터인 MIMIC-III (Medical Information Mart for Intensive Care) 를 전처리 및 그래프로 표현하고, GCT (Graph Convolutional Transformer) 모델을 학습시켰다. 학습 후, 어텐션 흐름 그래프를 시각화해서 모델의 예측에 대한 직관적인 설명을 제공한다.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.19 no.10
• /
• pp.633-639
• /
• 2018

High-resolution images using remote sensing (RS) is importance to secure for spatial classification depending on the characteristics of the complex and various factors that make up the river environment. The purpose of this study is to evaluate the accuracy of the classification results and to suggest the possibility of applying the high resolution hyperspectral images obtained by using the drone to perform spatial classification. Hyperspectral images obtained from study area were reduced the dimensionality with PCA and MNF transformation to remove effects of noise. Spatial classification was performed by supervised classifications such as MLC(Maximum Likelihood Classification), SVM(Support Vector Machine) and SAM(Spectral Angle Mapping). In overall, the highest classification accuracy was showed when the MLC supervised classification was used by MNF transformed image. However, it was confirmed that the misclassification was mainly found in the boundary of some classes including water body and the shadowing area. The results of this study can be used as basic data for remote sensing using drone and hyperspectral sensor, and it is expected that it can be applied to a wider range of river environments through the development of additional algorithms.

• Journal of Wetlands Research
• /
• v.7 no.2
• /
• pp.121-132
• /
• 2005

To classify the surface sedimentary facies using IKONOS image collected over Hwangdo tidal flat in Cheonsu Bay, the optical reflectance was compared for characterizing various sedimentary environments such as grain size, tidal channel pattern and area ratio of surface remnant water. The intertidal DEM (Digital Elevation Model) was generated by echo-sounder for analyzing the relationship between IKONOS image and sedimentary environments including topography. The boundary of the optical reflectance between mud-mixed facies and sand facies was distinct, and discrimination of the associated sandbar feature was also possible. The mud-mixed facies coupled with intricate tidal channels is confined to the relatively hi호 topography of Hwangdo tidal flat. The boundary between mud and mixed flat was indistinct in IKONOS optical reflectance but it would have a difference in the area ratio of surface remnant water. The dark area in the image represented the well developed sand facies having a lot of surface remnant water due to the relatively low surface topography. The overall accuracy of characterizing the surface sediment facies by maximum likelihood classification method was 86.2 %. These results demonstrate that high spatial resolution satellite imagery such as IKONOS coupled with knowledge of grain size, surface remnant water and tidal channel network can be effectively used to characterize the surface sedimentary facies (mud, mixed and sand) network of the tidal flat environments.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.51 no.3
• /
• pp.164-173
• /
• 2014

In underwater robotics, vision would be a key element for recognition in underwater environments. However, due to turbidity an underwater optical camera is rarely available. An underwater imaging sonar, as an alternative, delivers low quality sonar images which are not stable and accurate enough to find out natural objects by image processing. For this, artificial landmarks based on the characteristics of ultrasonic waves and their recognition method by a shape matrix transformation were proposed and were proven in Part 1. But, this is not working properly in undulating and dynamically noisy sea-bottom. To solve this, we propose a framework providing a selection phase of likelihood candidates, a selection phase for final candidates, recognition phase and tracking phase in sequence images, where a particle filter based selection mechanism to eliminate fake candidates and a mean shift based tracking algorithm are also proposed. All 4 steps are running in parallel and real-time processing. The proposed framework is flexible to add and to modify internal algorithms. A pool test and sea trial are carried out to prove the performance, and detail analysis of experimental results are done. Information is obtained from tracking phase such as relative distance, bearing will be expected to be used for control and navigation of underwater robots.

• Journal of Korean Society for Geospatial Information Science
• /
• v.18 no.4
• /
• pp.141-149
• /
• 2010

Three-dimensional geo-spatial information is important for the efficient use and management of the country and the three-dimensional expression and analysis of urban projects, such as urban plans devised by local governments and urban management. Thanks to the revitalization of the geo-spatial information service industry, it is now being variously used not only in public but also private areas. For the creation of high-guiltily three-dimensional geo-spatial information, emphasis should be placed on not only the quality of the source image and three-dimensional geo-spatial model but also the level of visualization, such as level of detail and texturing. However, in the case of existing three-dimensional geo-spatial information, its establishment process is complicated and its data are not updated frequently enough, as it uses ready-created digital maps. In addition, as it uses Ortho Images, the images exist Relief displacement. As a result, the visibility is low and the three-dimensional models of artificial features are simplified to reach LoD between 2 and 3, making the images look less realistic. Therefore, this paper, analyzed the quality of three-dimensional geo-spatial information created using the three-dimensional modeling technique were applied using Digital photogrammetry technique, using digital aerial photo images by an existing large-format digital camera and multi-looking camera. The analysis of the accuracy of visualization information of three-dimensional models showed that the source image alone, without other visualization information, secured the accuracy of 84% or more and that the establishment of three-dimensional spatial information carried out simultaneously with filming made it easier to gain the latest data. The analysis of the location accuracy of true Ortho images used in the work process showed that the location accuracy was better than the allowable horizontal position accuracy of 1:1,000 digital maps.

• Journal of the Institute of Electronics Engineers of Korea SD
• /
• v.39 no.11
• /
• pp.24-32
• /
• 2002

In this paper we implemented various image processing filtering using the format converter. This design method is based on realized the large processor-per-pixel array by integrated circuit technology. These two types of integrated structure are can be classify associative parallel processor and parallel process DRAM(or SRAM) cell. Layout pitch of one-bit-wide logic is identical memory cell pitch to array high density PEs in integrate structure. This format converter design has control path implementation efficiently, and can be utilize the high technology without complicated controller hardware. Sequence of array instruction are generated by host computer before process start, and instructions are saved on unit controller. Host computer is executed the pixel-parallel operation starting at saved instructions after processing start. As a result, we obtained three result that 1)simple smoothing suppresses higher spatial frequencies, reducing noise but also blurring edges, 2) a smoothing and segmentation process reduces noise while preserving sharp edges, and 3) median filtering, like smoothing and segmentation, may be applied to reduce image noise. Median filtering eliminates spikes while maintaining sharp edges and preserving monotonic variations in pixel values.

• Journal of Digital Contents Society
• /
• v.4 no.2
• /
• pp.211-221
• /
• 2003

The objective of this paper is performance analysis of the digital document image decimation algorithm which generates a value of decimated element by an average of a target pixel value and a value of neighbor intelligible element to adaptively reflect the merits of ZOD method and FOD method on the decimated image. First, a target pixel located at the center of sliding window is selected, then the gradient amplitudes of its right neighbor pixel and its lower neighbor pixel are calculated using first order derivative operator respectively. Secondly, each gradient amplitude is divided by the summation result of two gradient amplitudes to generate each local intelligible weight. Next, a value of neighbor intelligible element is obtained by adding a value of the right neighbor pixel times its local intelligible weight to a value of the lower neighbor pixel times its intelligible weight. The decimated image can be acquired by applying the process repetitively to all pixels in input image which generates the value of decimated element by calculating the average of the target pixel value and the value of neighbor intelligible element. In this paper, the performance comparison of proposed method and conventional methods in terms of subjective performance and hardware complexity is analyzed and the preferable approach for developing the decimation algorithm of the digital document image on the basis of this analysis result has been reviewed.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.22 no.4
• /
• pp.637-637
• /
• 1997

In this paper, format conversion filters are proposed, which have advantages in hardware implementation compared to the ones proposed in MPEG-4 Video Verification Model. since each coefficients of the proposed filters is constrained to have less than two non-zero digits in minimal signed digit representation, multiplication of input and the coefficient can be implemented by a single adder. As a result, the proposed filters have advantages in hardware complexity and speed, compared to the filters which are usually implemented by integer multiplier or carry save adders. Six kinds of filters are proposed in MPEG-4 Video Verification Model for size conversion of 2:1, 4:1, 5:3 and 5:6. We design 5 filters for the same purpose and compare the performance. The remaining one is very simple to implement. For comparing the filtering performance, we first compare the results of sine wave frequency conversion as an indirect but meaningful comparison. Second. We compute the PSNR of the images obtained from the proposed filters and the ones proposed by MPEG, with reference to the images obtained by using double precision arithmetic and high order filter. The results show that the performance of the proposed filters is almost the same as that of the filters proposed by MPEG. In conclusion, the peroformance of the proposed filters is comparable to that of the ones in MPEG-4, while requiring lower hardware complexity and providing high operating speed.

• Journal of Biomedical Engineering Research
• /
• v.19 no.4
• /
• pp.341-350
• /
• 1998

The accurate measurement of the femoral anteversion is important for the derotational osteotomy. To estimate femoral anteversion, following three major parameters are required; the neck axis, the long axis, and the knee axis. Conventional methods on the basis of 2D images are ambiguous to determine these major axes. As the femur has a complex 3 dimensional structure, the 3 dimensional model should be applied for accurate and reliable measurement of femoral anteversion. In this thesis, we model femur and define three parameters. The neck axis is defined from the femoral head and neck model. The long axis is determined from the cylindrical model of the femoral shaft. The knee axis is also determined from the model of femoral condyles. According to the definition of the femoral anteversion, the femoral anteversion is efficiently estimated from these models. 20 specimens were tested by the conventional 2D imaging method and 3D imaging method witch was developed by authors and the new 3D modeling method. The study provides accurate, fast and human factor free measurement for femoral anteversion.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.12 no.3
• /
• pp.179-185
• /
• 2019

This paper proposes a block-type classification based image binarization for the implementation of the low-power feature extraction algorithm. The proposed method can be implemented with threshold value re-use technique approach when the image divided into $64{\times}64$ macro blocks size and calculating the threshold value for each block type only once. The algorithm is validated based on quantitative results that only a threshold value change rate of up to 9% occurs within the same image/block type. Existing algorithms should compute the threshold value for 64 blocks when the macro block is divided by $64{\times}64$ on the basis of $512{\times}512$ images, but all suggestions can be made only once for best cases where the same block type is printed, and for the remaining 63 blocks, the adaptive threshold calculation can be reduced by only performing a block type classification process. The threshold calculation operation is performed five times when all block types occur, and only the block type separation process can be performed for the remaining 59 blocks, so 93% adaptive threshold calculation operation can be reduced.