
http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
회화의 조형 요소에 기반한 특징 추출 및 자기조직화지도를 이용한 시각화
Currently, art paintings are being utilized in various fields via digitalization using equipment such as the digital camera. This study suggests a method for extracting features from each art painting image, classifying them according to an objectified criterion, and visualizing them. Accordingly, a total of 62 features based on the formative elements of color, composition, and texture are extracted. By applying a statistical analysis on the extracted features, the tests of significance in accordance with the criterion are performed. The results will be used to analyze the significant differences between the art trends. Furthermore, self-organizing maps are used to create a visualization tool in order to classify the art paintings using the extracted features and to perform diversified analyses using the resulting feature map. By utilizing the features extracted from the experiment, the classification of art paintings was found to be feasible. It was also confirmed that the classification performance can be enhanced by using specific features selected from the tests of significance. Owing to visualization attained using this tool, the classifications of both art trends and works of artists were utilized for the analysis, thus suggesting a confluent utilization method for classifying a newly introduced type of art paintings.
본 논문에서는 경동맥 초음파 영상에서 내막-중막 두께(IMT ; Intima-Media Thickness)의 자동측정을 위한 초음파 영상 분석 시스템을 제안한다. 동맥경화 진단을 위한 중요한 지표인 IMT 측정에 대한 많은 연구들이 존재하지만, 모든 과정이 자동화되어 있지 않았다. 이런 수동 측정은 검사자의 주관적인 측정 위치 설정, 조작의 부정확성으로 추출 결과의 변이성이 발생하고 측정시간이 길어 비효율적이다. 따라서 본 논문에서는 경동맥 초음파 영상을 통해 IMT 측정을 위한 모든 과정을 자동화한 경동맥 초음파 영상 분석 시스템을 제안한다. 제안된 방법은 혈관 벽의 구조 및 경동맥의 연속성과 같은 본 논문에서 정의된 형태학적 특징을 사용하여 경동맥의 내막-중막 영역을 자동으로 검출하는 것이다. 이를 위해 우선 B-mode(밝기 방식, brightness mode) 초음파 영상에서 전처리과정을 통해 스펙클 잡음을 먼저 제거한 후 경동맥의 혈관 영역을 검출하기 위해 형태학적 특징을 강조한다. 이는 가우시안 블러, 엔즈-인 탐색 스트레칭, 자기 조직화 맵을 이용한 색상 양자화 및 영상의 형태학적 특징 강조 연산을 이용한다. 다음 경동맥 직선화 알고리즘을 수행하여 경동맥 초음파 영상에서 내막-중막 두께 측정 위치를 자동 검출하고 검출된 영역에서 IMT를 측정하는 방법을 제안한다. 제안한 방법의 IMT 측정 위치 자동 검출 및 측정 결과의 정확도를 평가하기 위해서 전문의가 자동 검출 결과 영상을 확인하였고, 상용 반자동 IMT 측정프로그램을 이용한 검출 영상과 비교 하였다. 이를 통해 제안된 의료 초음파 영상 분석 시스템이 경동맥의 형태학적 특징을 고려함으로써 내막-중막 두께 측정을 위한 영역을 자동 검출하고 측정에 뛰어난 성능을 보여주는 것을 확인하였다.
컨볼루션 신경망과 조건부 무작위장을 이용한 의미론적 영상 분할에 관한 연구
영상 분할 문제는 무인자동차, 보안 시스템, 의료, 공정 검사 등 광범위한 분야에서 사용되고 있으며 최근 인공지능 기술의 발전에 힘입어 비약적인 성능 향상과 기술에 대한 관심이 더욱 높아졌다. 일반적인 영상 분할은 주어진 영상을 의미 있는 객체별로 분할만 하는 것이지만 본 논문에서 다룰 의미론적 영상 분할 문제는 영상 분할 뿐만 아니라 각 객체가 무엇인지 까지 판별하는 문제이다. 의미론적 영상 분할 문제를 해결하기 위해 임계값을 이용한 이진화, 윤곽선을 검출하여 객체 간 경계를 찾는 방법, 후 처리 과정으로 조건부 무작위장을 이용하는 방법들이 연구되어 왔으며 최근에는 심화 학습(Deep learning) 중 하나인 CNN(Convolutional Neural Network)를 이용한 방법이 활발히 연구 중이다. 본 논문에서는 CNN 모델 중 하나인 ResNet을 기반으로 의미론적 영상 분할에 적합한 컨볼루션 신경망을 구성한다. 일반적인 컨볼루션 필터 대신 관심영역 확장을 위해 Dilated 컨볼루션 필터를 사용하며 주로 사용되는 손실 함수인 크로스 엔트로피와 영상 분할 정확도를 나타내는 IoU를 이용한 손실 함수를 사용하여 비교한다. 후처리 과정에서는 객체의 세밀한 픽셀 분할을 위한 완전 연결 조건부 무장위장을 사용한다. 실험 데이터로는 Pascal Voc2012를 사용하며 학습 데이터 증강을 위해 자르기, 반전, 회전 등의 방법으로 원본 데이터를 조작한 학습 데이터를 추가하였다. 실험 결과로는 ResNet을 기반으로 한 151계층 구조에 IoU를 이용한 함수를 사용한 실험에서 평균 IoU가 0.752로 가장 높은 결과를 얻었으며 계층을 더 깊게 하고 손실 함수를 달리하였을 때 좋은 결과를 얻을 수 있다는 것을 확인하였다.
외형 기반의 다중 특징 추출과 직접 매칭을 이용한 Person Re-identification 방법
This study suggests an improved method to add color name, HOG, and LBP to SDALF, one of the methods of person re-identification. Since SDALF utilizes color information-based features to match images, it is improved by adding color name which is robust color information to illumination change and viewpoint change, HOG as shape information and LBP as texture information. VIPeR Dataset and ETHZ Dataset were used for Data set. Experiment used single shot and multiple shot method. The CMC curve was used for performance evaluation. The performance of the proposed method is higher than that of SDALF.
다중 추정 기반의 객체 위치 확률을 이용한 초음파 영상의 담낭 검출
Objects must be separated and recognized to acquire information from an image. Although image processing techniques have been researching, it is still difficult to detect a specific object in an image. If the object has no consistent features such as color, size, shape, location, and contour, the existing object detecting methods cannot be used. Therefore, this study suggests an object location probability for detecting these objects using brightness features. It estimates the probability that an object can be located by using contrast enhancement and color quantization. Also, this study proposes a gallbladder detection method in ultrasonogram using object location probability. The gallbladder candidate is generated by using the object location probability. Support vector machine is used to determine which the extracted candidate is the gallbladder, and the post-processing step selects a final gallbladder candidate. The final gallbladder candidate is segmented by using modified grassfire algorithm, and whole processes are completed. In experiment, 98 ultrasonograms are used to evaluate the performance of the proposed method. The ultrasonograms have various shapes and sizes of gallbladders, and were used in actual in-hospital (tertiary hospital) diagnoses. Also, this study generates the ground truth to measure the accuracy of the detected gallbladder. Experimental results show that the object location probability accurately estimates the probability of the object being located and the gallbladder detection method can detect allbladder more accurately than the previous researches, and the detected results are more accurate.
정보손실이 적은 퍼지논리 기반의 동적 적응 이진화 방법
Image recognition techniques are one of frontier technologies, and its importance is recently getting higher. Binarization classifies pixels into black and white depending on the threshold to separate the image into objects and backgrounds. Therefore, it is very important preprocessing in image processing for detecting the objects. Binarization is divided into global binarization, adaptive binarization, and fuzzy binarization. Global binarizations set a threshold for a whole image, so if there is an object gets irregular illuminance or an object has similar pixel values with the threshold, then it is hard to describe objects. adaptive binarizations divide the image into M × N subimages, and set a threshold for each of subimages to binarize them. Therefore, adaptive binarizations good results for each subimages, but it is very unnatural result for entire image because there are boundaries between the subimages. Thus, fuzzy binarization set two thresholds to separate the objects and the backgrounds, but it has a limitation that is suitable only for specific images. In this paper, the goal of the proposed method is not only to complement the poor description of objects in global binarization and fuzzy binarization but also unnatural result in adaptive binarization. The proposed method uses dynamic adaptive binarization approach which is splits the image stage by stage. It classifies the image into obvious region and ambiguous region at every stage by using fuzzy logic. The obvious regions are binarized immediately but the ambiguous regions are split by fuzzy logic repeatedly. Experimental results show the proposed method describes objects in detail than the global binarization and gives more natural results than adaptive binarization, and the proposed method can be used more widely than the fuzzy binarization by resolving the limitation. Additionally, it binarized the images with less information loss than other binarizations.
VR 환경을 위한 파노라마 영상에서 사람 영역 제거 및 배경 복원
In this thesis, I propose a method to remove the human region and reconstruct the background from the panoramic image. Conventional methods developed on the basis of planar images can not be applied to panoramic images with different coordinate system characteristics. Therefore, the algorithm should be applied to the geometric characteristics of the panoramic image. The panoramic image can be represented by a spherical coordinate system that appears through latitude and longitude. In order to detect the HOG feature descriptor in the panoramic image, obtain a coordinate system whose distortion is corrected in cell units. then, the SVM algorithm is applied to the detected HOG feature descriptor to detect a human region. Removes the detected human region and restores the background. The empty area is restored to the background through an example-based inpainting method. When the inpainting is applied, find the coordinates taking distortion into account and compare the similarities. Bounding techniques are also applied to speed up inpainting. Experimental results show that the proposed method successfully performs in panoramic images.
배경 추정 기법과 파티클 필터를 이용한 객체 검출 및 추적 방법
A Object Detection and Tracking algorithm is proposed in this paper for security and surveillance systems that can be applied to real-time. This algorithm is a new and improved method after analysis that object detection using indifferential image and tracking points of object. In this paper, Gaussian Mixture Model is used to detect objects, and Particle Filter of statistical probability model was used to track the objects in sequence of images. Experimental results show that proposed method effectively performed well at Object detection and tracking in PETS 2009 Benchmark Database and gave high performance.
워터쉐드 알고리즘을 이용한 세라믹 항공갑판 영상에서의 결함 검출 방법
In this paper, we propose a method for detecting flaws in air deck images obtained from X-Ray prints by applying Watershed algorithm. Those proposed methods consist of two main stages. In the first stage, we analyze the profile of image by finding the direction which brightness of its background region constantly changes along. And then based on information from the profile, we eliminate the variation of its brightness using Least Squares method, and find the ROI that has a flaw. In the second stage, we split the ROI using Watershed algorithm, and thresholding method is applied for each region to detect the flaw. With our proposed method, its precision was 96.57% and its recall was 87.04%, and these results are 3.47%p and 5.24%p higher respectively, compared with previous strategy.