• Proceedings of the Korea Information Processing Society Conference
• /
• 2004.11a
• /
• pp.793-796
• /
• 2004

이 논문에서 제안한 ICA기반 피처추출 알고리즘은 다차원 영상에서 각 픽셀의 반사도 분광영역이 서로 다른 물체타입(목표피처와 배경피처)으로 이루어진 선형 혼합 분광영역으로 가정되는 픽셀에 대한 목표피처 탐지를 목적으로 한다. Landsat ETM+ 위성영상은 다차원 데이터구조로 이루어져 있으며, 영상에는 추출하고자하는 목표피처와 여러 종류의 배경피처들이 혼재한다. 이 논문에서는 목표피처(김양식장) 주변의 배경피처(갯뻘, 바닷물 등)들을 효과적으로 제거하기 위하여 목표피처의 픽셀 분광영역을 배경피처의 픽셀 분광영역으로 직교투영하게 된다. 픽셀내의 나머지 목표피처 분광영역의 양은 배경피처의 분광영역을 제거함으로써 추정하게 된다. 이 논문에서 제안한 ICA기반의 피처추출 방법의 우수성을 확인하기 위하여 Landsat ETM+ 위성영상에서 김양식장 피처를 추출하는데 적용하였다. 또한 피처추출 후 제거되지 않고 남아 있는 잡음(noise)정도와 피처추출 정확도 측면에서 전통적으로 가장 많이 사용되고 있는 최대우도 분류방법과 비교실험을 하였다. 결과적으로 이 논문에서 제안하는 방법이 목표피처 주변의 혼합분광영역에서 배경피처를 효과적으로 제거하여 추출하고자 하는 목표피처를 추출하는데 있어 우수한 탐지 성능을 보임을 알 수 있었다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2013.06a
• /
• pp.52-55
• /
• 2013

본 논문은 Boosted Random Ferns 기반의 회전 불변 얼굴 검출 방법을 제안한다. 기존 Random Ferns 의 경우 특징값을 추출할 때 임의로 선택한 두 픽셀의 밝기값 비교를 통하여 이진 특징값을 추출한다. 이 경우 해당 픽셀의 밝기값에 잡음이 포함되면 특징값이 부정확하게 추출되는 문제가 있다. 본 논문에서는 이러한 문제를 해결하기 위하여 임의로 두 블록을 선택하고 해당 블록내 밝기값의 평균을 비교하여 이진 특징값을 추출하였다. 또한 픽셀 위치를 임의로 선택하여 ferns 를 구성하였던 기존의 방법 대신 최고의 분류 성능을 가지는 fern 들을 이용하여 분류기를 구성하기 위해, AdaBoost 의 방법을 Random Ferns 에 맞게 변경하였다. Boosted Random Ferns 를 트리 구조의 cascade 노드에 방향과 각도에 따라 배치하여 연산 속도를 향상시키고 false-positive를 줄이는 효과를 보았다. CMU Rotated Face Database 를 사용하여 평가하였을 때, 기존 Random Ferns 는 false-positive 의 수가 57 개 일 때 66%의 검출률을 보인 반면, Boosted Random Ferns 는 false-positive 의 수가 45 개 일 때 88%의 검출률을 보였다.

• Journal of the Korea Society of Computer and Information
• /
• v.11 no.2 s.40
• /
• pp.25-33
• /
• 2006

We propose a method that automatically classifies the images into the object and non-object images. An object image is the image with object(s). An object in an image is defined as a set of regions that lie around center of the image and have significant color distribution against the other surround (or background) regions. We define four measures based on the characteristics of an object to classify the images. The center significance is calculated from the difference in color distribution between the center area and its surrounding region. Second measure is the variance of significantly correlated colors in the image plane. Significantly correlated colors are first defined as the colors of two adjacent pixels that appear more frequently around center of an image rather than at the background of the image. Third one is edge strength at the boundary of candidate for the object. By the way, it is computationally expensive to extract third value because central objects are extracted. So, we define fourth measure which is similar with third measure in characteristic. Fourth one can be calculated more fast but show less accuracy than third one. To classify the images we combine each measure by training the neural network and SYM. We compare classification accuracies of these two classifiers.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2015.11a
• /
• pp.134-136
• /
• 2015

영상 역 하프토닝은 입력된 하프톤 영상으로부터 그레이 영상을 복원시키는 것으로, 하프톤 영상으로 처리하지 못하는 다양한 영상처리를 가능하게 해주는 방법이다. 기존의 참조표를 이용한 역 하프토닝 방법은 다양한 하프톤 영상과 원본 그레이 영상으로부터 추출한 정보를 이용해 입력 영상을 복원시키는데, 본 논문에서는 이를 바탕으로 하여 영상의 질을 전반적으로 향상시킬 수 있는 국부적인 이진 패턴 기반 참조표를 이용한 영상 역 하프토닝 방법을 제안한다. 먼저 참조표를 이용한 역하프토닝 방법을 이용해 영상을 복원한 후 각 픽셀에서의 국부 이진패턴을 계산하여 각 픽셀 값을 패턴에 따라 분류한다. 분류된 패턴 정보에 따라 국부 이진 패턴 기반 참조표를 생성하고 이를 통해 입력 하프톤 영상에 대한 역 하프토닝을 수행한다. 실험 결과는 제안하는 알고리즘이 오류 확산법에 의해 변환된 하프톤 이미지를 역 하프토닝 했을 때, 기존의 역 하프토닝 방법에 비해 더 나은 PSNR을 달성하는 것을 보인다.

• Journal of Korean Society of Forest Science
• /
• v.100 no.4
• /
• pp.540-547
• /
• 2011

This study was conducted to compare classification accuracy by tree species using QuickBird imagery for pixel- and segment-based classifications that have been mostly applied to classify land covers. A total of 398 points was used as training and reference data. Based on this points, the points were classified into fourteen land cover classes: four coniferous and seven deciduous tree species in forest classes, and three non-forested classes. In pixel-based classification, three images obtained by using raw spectral values, three tasseled indices, and three components from principal component analysis were produced. For the both classification processes, the maximum likelihood method was applied. In the pixel-based classification, it was resulted that the classification accuracy with raw spectral values was better than those by the other band combinations. As resulted that, the segment-based classification with a scale factor of 50% provided the most accurate classification (overall accuracy:76% and ${\hat{k}}$ value:0.74) compared to the other scale factors and pixel-based classification.

• Korean Journal of Remote Sensing
• /
• v.34 no.2_2
• /
• pp.339-349
• /
• 2018

The massive green tide occurred every summer in the Yellow Sea since 2008, and many studies are being actively conducted to estimate the coverage of green tide through analysis of satellite imagery. However, there is no satellite images selection criterion for accurate coverage calculation of green tide. Therefore, this study aimed to find a suitable satellite image from for the comparison of the green tide coverage according to the spatial resolution of satellite image. In this study, Landsat ETM+, MODIS and GOCI images were used to coverage estimation and its spatial resolution is 30, 250 and 500 m, respectively. Green tide pixels were classified based on the NDVI algorithm, the difference of the green tide coverage was compared with threshold value. In addition, we estimate the proportion of the green tide in one pixel through the Linear Spectral Unmixing (LSU) method, and the effect of the difference of green tide ratio on the coverage calculation were evaluated. The result of green tide coverage from the calculation of the NDVI value, coverage of green tide usually overestimate with decreasing spatial resolution, maximum difference shows 1.5 times. In addition, most of the pixels were included in the group with less than 0.1 (10%) LSU value, and above 0.5 (50%) LSU value accounted for about 2% in all of three images. Even though classified as green tide from the NDVI result, it is considered to be overestimated because it is regarded as the same coverage even if green tide is not 100% filled in one pixel. Mixed-pixel problem seems to be more severe with spatial resolution decreases.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.6
• /
• pp.117-123
• /
• 2009

In this paper, An algorithm which converts 2D video frames into 3D video frames of parallel looking stereo camea is proposed. The proposed algorithm finds the disparity information between two consecutive video frames and generates 3D video frames from the obtained disparity maps. The disparity information is obtained from the modified iterative convergence algorithm. The method of generating 3D video frames from the disparity information is also proposed. The proposed algorithm uses coherence method which overcomes the video pattern based algorithms.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.10 no.3
• /
• pp.252-259
• /
• 2000

In this paper, we propose a Fuzzy Classifier System(FCS) to find a set of fuzzy rules which can carry out the edge detection of a image. The FCS is based on the fuzzy logic system combined with machine learning. Therefore the antecedent and consequent of a classifier in FCS are the same as those of a fuzzy rule. There are two different approaches, Michigan and Pittsburgh approaches, to acquire appropriate fuzzy rules by evolutionary computation. In this paper, we use the Michigan style in which a single fuzzy if-then rule is coded as an individual. Also the FCS employs the Genetic Algorithms to generate new rules and modify rules when performance of the system needs to be improved. The proposed method is evaluated by applying it to the edge detection of a gray-level image that is a pre-processing step of the computer vision. the differences of average gray-level of the each vertical/horizontal arrays of neighborhood pixels are represented into fuzzy sets, and then the center pixel is decided whether it is edge pixel or not using fuzzy if-then rules. We compare the resulting image with a conventional edge image obtained by the other edge detection method such as Sobel edge detection.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.48 no.5
• /
• pp.16-23
• /
• 2011

Muzzle flash based on gunnery is the target that has huge energy. So, gunnery target in a long range over xx km is distinguishable in the IR(infrared) images, on the other hand, is not distinguishable in the CCD images. In this paper, we propose the classification method of gunnery targets in a infrared images and in a long range. The energy from gunnery have an effect on varous pixel values in infrared images as a property of infrared image sensor, distance, and atmosphere, etc. For this reason, it is difficult to classify gunnery targets using pixel values in infrared images. In proposed method, we take the profile of pixel values using high performance infrared sensor, and classify gunnery targets using modeling GMM and shape of profile. we experiment on the proposed method with infrared images in the ground and aviation. In experimental result, the proposed method provides about 93% classification rate.

• Journal of the Korean association of regional geographers
• /
• v.12 no.4
• /
• pp.496-507
• /
• 2006

A hybrid approach integrating both high-resolution and hyperspectral data sets was used to map vegetation cover of a coastal lowland area in the Hawaii Volcanoes National Park. Three common grass species (broomsedge, natal redtop, and pili) and other non-grass species, primarily shrubs, were focused in the study. A 3-step, hybrid approach, combining an unsupervised and a supervised classification schemes, was applied to the vegetation mapping. First, the IKONOS 1-m high-resolution data were classified to create a binary image (vegetated vs. non--vegetated) and converted to 20-meter resolution percent cover vegetation data to match AVIRIS data pixels. Second, the minimum noise fraction (MNF) transformation was used to extract a coherent dimensionality from the original AVIRIS data. Since the grasses and shubs were sparsely distributed and most image pixels were intermingled with lava surfaces, the reflectance component of lava was filtered out with a binary fractional cover analysis assuming that tile total reflectance of a pixel was a linear combination of the reflectance spectra of vegetation and the lava surface. Finally, a supervised approach was used to classify the plant species based on tile maximum likelihood algorithm.