머신러닝 기법을 적용한 CS-RANSAC 알고리즘

고승현 ( Seunghyun Ko ),윤의녕 ( Ui-nyoung Yoon ),주마백 ( Jumabek Alikhanov ),조근식 ( Geun-sik Jo ) 한국정보처리학회 2016 한국정보처리학회 학술대회논문집 Vol.23 No.2

증강현실에서 영상과 증강된 콘텐츠 간의 이질감을 줄이기 위해서 정확한 호모그래피 행렬을 추정해야 하며, 정확한 호모그래피 행렬을 추정할때 RANSAC 알고리즘이 널리 사용된다. 그러나 RANSAC 알고리즘은 랜덤 샘플링 과정을 반복적으로 거치기 때문에 불필요한 연산 과정이 발생하고 이로 인해 알고리즘의 효율이 저하된다. 이러한 단점을 극복하기 위해 DCS-RANSAC 알고리즘이 제안 되었다. 제안된 DCS-RANSAC 알고리즘은 이미지를 특징점 분포 패턴에 따라 그룹으로 분류하고 각 그룹에 제약조건 문제를 적용하여 불필요한 연산 과정을 줄이고 정확도를 향상시킨 알고리즘이다. 그러나 DCS-RANSAC 알고리즘에서 사용된 이미지 그룹 데이터는 수동적인 방법을 통해 직관적으로 분류되어 있지만 특징점 분포 패턴이 다양하지 않아 분류시 정확도가 저하되는 경우가 있다. 위의 문제점을 해결하기 위해 본 논문에서는 머신러닝 기법을 통해 이미지들을 자동으로 분류하고 각 그룹마다 각기 다른 제약조건을 적용하는 MCS-RANSAC 알고리즘을 제안한다. 제안하는 알고리즘은 머신러닝 기법을 사용하여 전처리 단계에서 이미지를 분류하고 분류된 이미지에 제약조건을 적용시켜 알고리즘의 처리시간을 줄이고 정확도를 향상시켰다. 실험 결과 본 논문에서 제안하는 MCS-RANSAC은 DCS-RANSAC 알고리즘에 비해 수행시간이 약 6% 단축되었고 호모그래피 오차율은 약 15% 줄어들었으며 참정보 비율은 2.8% 증가한 것으로 확인되었다.

K-Means 클러스터링을 적용한 향상된 CS-RANSAC 알고리즘

고승현 ( Seunghyun Ko ),윤의녕 ( Ui-nyoung Yoon ),( Jumabek Alikhanov ),조근식 ( Geun-sik Jo ) 한국정보처리학회 2017 정보처리학회논문지. 소프트웨어 및 데이터 공학 Vol.6 No.6

이미지를 기반으로 하는 증강현실 시스템에서 가상의 객체를 실제 영상에 저작할 때 생기는 이질감을 줄이기 위해서는 실제 영상에 저작된 가상객체의 방향과 위치에 대해 정확하게 추정을 해야 하며, 이때 호모그래피를 사용한다. 호모그래피를 추정하기 위해서는 SURF와 같은 특징점을 추출하고 추출된 특징점들을 통해 호모그래피 행렬을 추정한다. 호모그래피 행렬의 추정을 위해서 RANSAC 알고리즘이 주로 사용되고 있으며, 특히 RANSAC에 제약 조건 만족 문제(Constraint Satisfaction Problem)와 여기에 사용되는 제약조건을 동적으로 적용하여 속도와 정확도를 높인 DCS-RANSAC 알고리즘이 연구되었다. DCS-RANSAC 알고리즘에서 사용된 이미지 그룹 데이터는 수동적인 방법을 통해 직관적으로 분류되어 있지만 특징점 분포 패턴이 다양하지 않고, 이미지들을 정확하게 분류하기가 어려워서 이로 인해 알고리즘의 성능이 저하되는 경우가 있다. 따라서 본 논문에서는 K-means 클러스터링을 적용하여 이미지들을 자동으로 분류하고 각 이미지 그룹마다 각기 다른 제약조건을 적용하는 KCS-RANSAC 알고리즘을 제안한다. 제안하는 알고리즘은 머신러닝 기법인 K-means 클러스터링을 사용하여 전처리 단계에서 이미지를 특징점 분포 패턴에 따라 자동으로 분류하고, 분류된 이미지에 제약조건을 적용하여 알고리즘의 속도와 정확도를 향상시켰다. 실험결과 본 논문에서 제안하는 KCS-RANSAC이 DCS-RANSAC 알고리즘에 비해 수행시간이 약 15% 단축되었고, 오차율은 약 35% 줄어들었으며, 참정보 비율은 약 14% 증가되었다. Estimating the correct pose of augmented objects on the real camera view efficiently is one of the most important questions in image tracking area. In computer vision, Homography is used for camera pose estimation in augmented reality system with markerless. To estimating Homography, several algorithm like SURF features which extracted from images are used. Based on extracted features, Homography is estimated. For this purpose, RANSAC algorithm is well used to estimate homography and DCS-RANSAC algorithm is researched which apply constraints dynamically based on Constraint Satisfaction Problem to improve performance. In DCS-RANSAC, however, the dataset is based on pattern of feature distribution of images manually, so this algorithm cannot classify the input image, pattern of feature distribution is not recognized in DCS-RANSAC algorithm, which lead to reduce it`s performance. To improve this problem, we suggest the KCS-RANSAC algorithm using K-means clustering in CS-RANSAC to cluster the images automatically based on pattern of feature distribution and apply constraints to each image groups. The suggested algorithm cluster the images automatically and apply the constraints to each clustered image groups. The experiment result shows that our KCS-RANSAC algorithm outperformed the DCS-RANSAC algorithm in terms of speed, accuracy, and inlier rate.

Automatic Transformation of Korean Fonts using Unbalanced U-net and Generative Adversarial Networks

Pangjia(방가),Seunghyun Ko(고승현),Yang Fang(방양),Geun-sik Jo(조근식) Korean Institute of Information Scientists and Eng 2019 정보과학회논문지 Vol.46 No.1

In this paper, we study the typography transfer problem: transferring a source font, to an analog font with a specified style. To solve the typography transfer problem, we treat the problem as an image-to-image translation problem, and propose an unbalanced u-net architecture based on Generative Adversarial Network(GAN). Unlike traditional balanced u-net architecture, architecture we proposed consists of two subnets: (1) an unbalanced u-net is responsible for transferring specified fonts style to another, while maintaining semantic and structure information; (2) an adversarial net. Our model uses a compound loss function that includes a L1 loss, a constant loss, and a binary GAN loss to facilitate generating desired target fonts. Experiments demonstrate that our proposed network leads to more stable training loss, with faster convergence speed in cheat loss, and avoids falling into a degradation problem in generating loss than balanced u-net.

다중 신경망 레이어에서 특징점을 선택하기 위한 전이 학습 기반의 AdaBoost 기법

주마백 ( Jumabek Alikhanov ),가명현 ( Myeong Hyeon Ga ),고승현 ( Seunghyun Ko ),조근식 ( Geun-sik Jo ) 한국정보처리학회 2016 한국정보처리학회 학술대회논문집 Vol.23 No.1

Convolutional Networks (ConvNets) are powerful models that learn hierarchies of visual features, which could also be used to obtain image representations for transfer learning. The basic pipeline for transfer learning is to first train a ConvNet on a large dataset (source task) and then use feed-forward units activation of the trained ConvNet as image representation for smaller datasets (target task). Our key contribution is to demonstrate superior performance of multiple ConvNet layer features over single ConvNet layer features. Combining multiple ConvNet layer features will result in more complex feature space with some features being repetitive. This requires some form of feature selection. We use AdaBoost with single stumps to implicitly select only distinct features that are useful towards classification from concatenated ConvNet features. Experimental results show that using multiple ConvNet layer activation features instead of single ConvNet layer features consistently will produce superior performance. Improvements becomes significant as we increase the distance between source task and the target task.

모터다이나모를 이용한 전기자전거 주행성능 평가에 관한 연구

강대경(Daekyung Kang),고승현(Seunghyun Ko),김정훈(Jeunghoon Kim),서상근(Sangguen Seo),이상배(Sangbae Lee),김민수(Minsoo Kim) 한국자동차공학회 2012 한국자동차공학회 학술대회 및 전시회 Vol.2012 No.11

We have studied about drive performance of E-bike such as energy consumption efficiency and maximum speed. These measurements have been performed by testing the drive motor on motor dynamometer. Road load occurred during drive are applied to drive motor as braking force of powder brake in motor dynamometer. Through experimental results, we have found that energy consumption and maximum speed of E-bike is about 0.15 km/Wh, 23.0 km/h, respectively. We expect that this test method for electric vehicle including E-bike will be necessary to test and evaluate performance and reliability of vehicles in case of severe and time-continuous environments.

Backbone Network for Object Detection with Multiple Dilated Convolutions and Feature Summation

Vani Natalia Kuntjono(바니 나탈리아 쿤트조노),Seunghyun Ko(고승현),Yang Fang(방양),Geunsik Jo(조근식) Korean Institute of Information Scientists and Eng 2018 정보과학회논문지 Vol.45 No.8

The advancement of CNN leads to the trend of using very deep convolutional neural network which contains more than 100 layers not only for object detection, but also for image segmentation and object classification. However, deep CNN requires lots of resources, and so is not suitable for people who have limited resources or real time requirements. In this paper, we propose a new backbone network for object detection with multiple dilated convolutions and feature summation. Feature summation enables easier flow of gradients and minimizes loss of spatial information that is caused by convolving. By using multiple dilated convolution, we can widen the receptive field of individual neurons without adding more parameters. Furthermore, by using a shallow neural network as a backbone network, our network can be trained and used in an environment with limited resources and without pre-training it in ImageNet dataset. Experiments demonstrate we achieved 71% and 38.2% of accuracy on Pascal VOC and MS COCO dataset, respectively.

Transfer Learning using Multiple ConvNet Layers Activation Features with Principal Component Analysis for Image Classification

Batkhuu Byambajav(바트후 뱜바자브),Jumabek Alikhanov(주마벡 알리하노브),Yang Fang(팡양),Seunghyun Ko(고승현),Geun Sik Jo(조근식) 한국지능정보시스템학회 2018 지능정보연구 Vol.24 No.1

Convolutional Neural Network (ConvNet) is one class of the powerful Deep Neural Network that can analyze and learn hierarchies of visual features. Originally, first neural network (Neocognitron) was introduced in the 80s. At that time, the neural network was not broadly used in both industry and academic field by cause of large-scale dataset shortage and low computational power. However, after a few decades later in 2012, Krizhevsky made a breakthrough on ILSVRC-12 visual recognition competition using Convolutional Neural Network. That breakthrough revived people interest in the neural network. The success of Convolutional Neural Network is achieved with two main factors. First of them is the emergence of advanced hardware (GPUs) for sufficient parallel computation. Second is the availability of large-scale datasets such as ImageNet (ILSVRC) dataset for training. Unfortunately, many new domains are bottlenecked by these factors. For most domains, it is difficult and requires lots of effort to gather large-scale dataset to train a ConvNet. Moreover, even if we have a large-scale dataset, training ConvNet from scratch is required expensive resource and time-consuming. These two obstacles can be solved by using transfer learning. Transfer learning is a method for transferring the knowledge from a source domain to new domain. There are two major Transfer learning cases. First one is ConvNet as fixed feature extractor, and the second one is Fine-tune the ConvNet on a new dataset. In the first case, using pre-trained ConvNet (such as on ImageNet) to compute feed-forward activations of the image into the ConvNet and extract activation features from specific layers. In the second case, replacing and retraining the ConvNet classifier on the new dataset, then fine-tune the weights of the pre-trained network with the backpropagation. In this paper, we focus on using multiple ConvNet layers as a fixed feature extractor only. However, applying features with high dimensional complexity that is directly extracted from multiple ConvNet layers is still a challenging problem. We observe that features extracted from multiple ConvNet layers address the different characteristics of the image which means better representation could be obtained by finding the optimal combination of multiple ConvNet layers. Based on that observation, we propose to employ multiple ConvNet layer representations for transfer learning instead of a single ConvNet layer representation. Overall, our primary pipeline has three steps. Firstly, images from target task are given as input to ConvNet, then that image will be feed-forwarded into pre-trained AlexNet, and the activation features from three fully connected convolutional layers are extracted. Secondly, activation features of three ConvNet layers are concatenated to obtain multiple ConvNet layers representation because it will gain more information about an image. When three fully connected layer features concatenated, the occurring image representation would have 9192 (4096+4096+1000) dimension features. However, features extracted from multiple ConvNet layers are redundant and noisy since they are extracted from the same ConvNet. Thus, a third step, we will use Principal Component Analysis (PCA) to select salient features before the training phase. When salient features are obtained, the classifier can classify image more accurately, and the performance of transfer learning can be improved. To evaluate proposed method, experiments are conducted in three standard datasets (Caltech-256, VOC07, and SUN397) to compare multiple ConvNet layer representations against single ConvNet layer representation by using PCA for feature selection and dimension reduction. Our experiments demonstrated the importance of feature selection for multiple ConvNet layer representation. Moreover, our proposed approach achieved 75.6% accuracy compared to 73.9% accuracy achieved by FC7 layer on the Caltech-256 dataset, 73.1% accuracy compared to

특징 연결과 깊이별 분리 컨볼루션을 이용한 효율적인 얼굴 감정인식 CNN

이명오(Myeong Oh Lee),윤의녕(Ui Nyoung Yoon),고승현(Seunghyun Ko),조근식(Geun-Sik Jo) 한국정보과학회 2019 한국정보과학회 학술발표논문집 Vol.2019 No.6