http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
SAM 미세조정을 이용한 대장 내시경 용종이미지 분할 성능분석 및 성능향상 연구
송무건,신영학 한국콘텐츠학회 2024 한국콘텐츠학회논문지 Vol.24 No.5
Segment Anything Model (SAM)은 광범위한 대용량 일반이미지 데이터셋으로 학습된 이미지 분할 모델이다. 본 연구에서는 대규모 이미지 분할 모델이 소규모 데이터를 이용하는 특정 의료 영상 분야인 대장 내시경 용종분할에서 충분한 성능을 보이는지 확인하고, 미세조정을 통한 성능향상 방법에 대해 연구하였다. 세 가지 학습 방법을 제안하였는데, 첫째는 Generative Adversarial Networks (GAN)기반의 실제와 유사한 용종이미지 생성을 통해 생성된 이미지 데이터를 추가하여 학습데이터를 확장하는 것, 둘째는 copy-pasting 방법을 이용하여 기존 이미지 데이터를 합성하여 새로운 이미지 데이터를 만들고, 이를 학습데이터에 추가하는 것, 셋째는 앞서 사용한 생성된 이미지 데이터와 copy-pasting로 만든 이미지 데이터를 이용하여 SAM을 사전학습하고, 모델을 기존의 학습데이터로 미세조정 하는 것이다. 이 세 가지 방법을 통해 SAM의 성능을 향상하려는 시도가 이루어졌으며, 기존 모델에 비해 Dice Similarity Coefficient (DSC)와 Mean Intersection over Union (mIoU) 성능향상을 확인하였다. Segment Anything Model (SAM) is an image segmentation model trained on a large-scale, general image dataset. In this study, we investigated whether such a large-scale image segmentation model can perform adequately in the specific medical imaging field of colonoscopy polyp segmentation using small-scale data, and we explored methods to enhance performance through fine-tuning. Three training methods were proposed. First, expanding the training dataset by incorporating generated images through Generative Adversarial Networks (GAN), aiming for realistic polyp image synthesis. Second, synthesizing new image data by merging existing images through copy-pasting, and adding them to the training data. Third, pre-training SAM with the generated and copy-pasted images, followed by fine-tuning the model with the original training data. These three methods were an attempt to improve the performance of SAM, and we observed enhancements in the Dice Similarity Coefficient (DSC) and Mean Intersection over Union (mIoU) compared to the original model.
송무건,백재용,신관수,유송민 한국공작기계학회 2001 한국공작기계학회 춘계학술대회논문집 Vol.2001 No.-
In this study a vision system with image processing method have been introduced to find the edge radius of curvature. It was applied to inspect the edge quality of the de burring process product with brush grinding. Size of data was found to be critical in calculating the radius of curvature. Results using laser measurement system were compared.
송무건,유송민 한국공작기계학회 2000 한국공작기계학회 추계학술대회논문집 Vol.2000 No.-
Surface roughness measurement system with capacitance type gap sensor. Tentative result from the calibration measurement showed the potential applicability of the sensor to the processed specimen. In order to test the sensitivity of the measurement system, several parameters including valley depth, width of the specimen have been changed. Effect of the charge area between sensor and specimen surface has been also analyzed.
실차 도로이력 측정시험 (RLDA)과 해석(CAE)을 이용한 차체 쇼타워 (Shock Tower) 하중 결과에 대한 상관 관계에 대한 연구
송무건(Mu Keon Song),한교진(Kyo Jin Han) 한국자동차공학회 2010 한국자동차공학회 학술대회 및 전시회 Vol.2010 No.11
For RLDA & CAE correlation, in the refinement of CAE models to accord with test results of the physical RLDA is a field in the today automotive industries. The accuracy result of RLDA & CAE is the biggest issue of correlation. It is therefore of importance that correlation can be set and performed to provide the required test quality for the various road and conditions. In particular, fatigue or durability is traditionally a test based activity?especially RLDA, but now CAE becomes one of important tool for development of vehicle. To correlate RLDA & CAE results, they have been preformed and analyzed on special mode - Max Pothole. The load of the vehicle on Shock Tower is investigated by using experiments based on RLDA. Also the well-correlated model is very necessary for the feasible evaluations of structural performances. To perform and process the correlation analysis between RLDA and CAE, the ADAMS model updating techniques were developed and implemented into the existing ADAMS Dynamics software and optimization design. This paper gives a design reference for the sensor development and technique. All of activity is performed and simulated in a GMDAT Technical Center.
Mu Keon Song(송무건),Seung Jun Lee(이승준),Kyo Jin Han(한교진) 한국자동차공학회 2007 한국자동차공학회 춘 추계 학술대회 논문집 Vol.- No.-
This paper describes the development of precision Multi-Component-Circle-load cell with using strain-gauge in road load data acquisition (RLDA) for vehicle engine mount. Engine mount system is so important that it shows performance of vehicle. Engine mounts properly locate the engine in the chassis, and are an important factor in how smoothly a vehicle operates. Engine mounts are never replaced in the service life of a car or truck. To predict life in laboratory, acquired data is needed. In automotive industry, usually the data is acquired on proving ground and real road. This process is called Road Load Data Acquisition (RLDA). In RLDA, loadcell design must be important for sampling and collecting data at proving ground and real road. The data is provided to lab engineers. Often mis-desinged loadcell can occur to acquire wrong data There are other factors such as linearity, hysteresis, zero drift, temperature compensation, elongation limit, dynamic response, stability, effect of moisture and humidity. It also makes wrong results in math and analysis. This paper focuses on crosstalk of load cells, which are self-manufactured and conventional. The conventional load cell is manufactured by Michigan Scientific. The self-manufactured load cell is designed by GMDAT. Both of them are compared crosstalk ratio by road load data from engine mount.
백재용,송무건,유송민 한국공작기계학회 2001 한국공작기계학회 춘계학술대회논문집 Vol.2001 No.-
An image processing method was applied to characterize a shape of the flexible grinding disk. A disk surface image was taken by CCD camera. Depth of cut was changed to be 2 and 4mm. Circles marked on the disk were captured to extract the key features of the deflection. Notable correlation has been observed between the intervals and the process conditions. Same methodology has been applied to check the symmetry of the human face. Tentative results revealed that symmetry could be checked using the filtered face image.
차민건(MinGeon Cha),송훈(Hun Song),송무건(MuGeon Song),신영학(Younghak Shin) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
With the development of IT technology, the number of buyers who purchase directly from overseas shopping malls is increasing. However, ordinary users who are not fluent in foreign languages have difficulties. In this study, in order to solve this problem, various deep learning techniques were used to make the English that on the web image appear naturally translated into Korean. Using OCR, image inpainting, and translation API, which are sentence extraction technologies, the English on the web page was translated into Korean so that the font size and background looked natural.