• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.10a
• /
• pp.946-949
• /
• 2019

한 벌의 한글 글자체를 만드는데 일반적으로 많은 제작 비용과 시간이 소요된다. 따라서 폰트 제작의 어려움을 덜기 위해, 사용자가 대표 글자들을 입력하면 그 글자들의 디자인 특성을 딥러닝 기술을 이용하여 학습한 모델이 나머지 글자들을 자동 생성해주는 시스템 구축한다면 폰트 제작이 훨씬 용이해질 뿐만 아니라 저작권 문제로부터 자유로워질 것이다. 이와 관련된 선행연구를 실행하고 분석해 본 결과 데이터 전처리 과정에서 글자가 잘리거나 크기가 맞지 않아 제대로 된 데이터셋이 구축되지 않는 문제가 있음을 발견하였다. 본 논문에서는 이러한 문제를 해결하기 위해 템플릿에서 자동적으로 글자영역을 추출하고 이미지를 보정하는 전처리 과정과 함께 기존 모델에서 새로운 필터를 추가하여 학습 성능을 높이는 방법을 제안한다. 이를 통해 기존 연구에서 측정된 손실값을 낮춘 결과를 확인했으며 결과적으로 실제 글자체와 더욱 유사한 사용자 맞춤형 글자체를 제공할 수 있을 것이다.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.54 no.1
• /
• pp.70-79
• /
• 2017

This study proposes a method to quickly and accurately acquire distance information on direction signs. The proposed method is composed of the recognition of the sign, pre-processing to facilitate the acquisition of the road sign distance, and the acquisition of the distance data. The road sign recognition uses color detection including gamma correction in order to mitigate various noise issues. In order to facilitate the acquisition of distance data, this study applied tilt correction using linear factors, and resolution correction using Fourier transform. To acquire the distance data, morphological operation was used to highlight the area, along with labeling and template matching. By acquiring the distance information on the direction sign through such a processes, the proposed system can be output the distance remaining to the next junction. As a result, when the proposed method is applied to system it can process the data in real-time using the fast calculation speed, average speed was shown to be 0.46 second per frame, with accuracy of 0.65 in similarity value.

• Journal of KIISE:Software and Applications
• /
• v.34 no.12
• /
• pp.1075-1082
• /
• 2007

In this paper, we propose the method of multiple human action recognition on video clip. For being invariant to the change of speed or size of actions, Spatial-Temporal Pyramid method is applied. Proposed method can minimize the complexity of the procedures owing to select Motion Gradient Histogram (MGH) based on statistical approach for action representation feature. For multiple action detection, Motion Energy Image (MEI) of binary frame difference accumulations is adapted and then we detect each action of which area is represented by MGH. The action MGH should be compared with pre-learning MGH having pyramid method. As a result, recognition can be done by the analyze between action MGH and pre-learning MGH. Ten video clips are used for evaluating the proposed method. We have various experiments such as mono action, multiple action, speed and site scale-changes, comparison with previous method. As a result, we can see that proposed method is simple and efficient to recognize multiple human action with stale variations.

• Journal of the Institute of Convergence Signal Processing
• /
• v.14 no.3
• /
• pp.162-168
• /
• 2013

This paper examines the implementation of automatic compensation modules for digital camera image when a person is closing his/her eyes. The modules detect the face and eye region and then recognize the eye state. If the image is taken when a person is closing his/her eyes, the function corrects the eye and produces the image by using the most satisfactory image of the eye state among the past frames stored in the buffer. In order to recognize the face and eye precisely, the pre-process of image correction is carried out using SURF algorithm and Homography method. For the detection of face and eye region, Haar-like feature algorithm is used. To decide whether the eye is open or not, similarity comparison method is used along with template matching of the eye region. The modules are tested in various facial environments and confirmed to effectively correct the images containing faces.

• Korean Journal of Remote Sensing
• /
• v.35 no.3
• /
• pp.457-470
• /
• 2019

Automatic Target Recognition (ATR) using Synthetic Aperture Radar (SAR) has been attracted attention in the fields of surveillance, reconnaissance, and national security due to its advantage of all-weather and day-and-night imaging capabilities. However, there have been some difficulties in automatically identifying targets in real situation due to various observational and environmental conditions. In this paper, ATR problems in Extended Operating Conditions (EOC) were investigated. In particular, we considered partial occlusions of the target (10% to 50%) and differences in the depression angle between training ($17^{\circ}$) and test data ($30^{\circ}$ and $45^{\circ}$). To simulate various occlusion conditions, SARBake algorithm was applied to Moving and Stationary Target Acquisition and Recognition (MSTAR) images. The ATR accuracies were evaluated by using the template matching and Adaboost algorithms. Experimental results on the depression angle showed that the target identification rate of the two algorithms decreased by more than 30% from the depression angle of $45^{\circ}$ to $30^{\circ}$. The accuracy of template matching was about 75.88% while Adaboost showed better results with an accuracy of about 86.80%. In the case of partial occlusion, the accuracy of template matching decreased significantly even in the slight occlusion (from 95.77% under no occlusion to 52.69% under 10% occlusion). The Adaboost algorithm showed better performance with an accuracy of 85.16% in no occlusion condition and 68.48% in 10% occlusion condition. Even in the 50% occlusion condition, the Adaboost provided an accuracy of 52.48%, which was much higher than the template matching (less than 30% under 50% occlusion).

• KIPS Transactions on Software and Data Engineering
• /
• v.9 no.5
• /
• pp.169-176
• /
• 2020

Recently, deep neural networks (DNNs) have brought revolutions to many mobile and embedded devices by providing human-level machine intelligence for various applications. However, high inference accuracy of such DNNs comes at high computational costs, and, hence, there have been significant efforts to reduce computational overheads of DNNs either by compressing off-the-shelf models or by designing a new small footprint DNN architecture tailored to resource constrained devices. One notable recent paradigm in designing small footprint DNN models is sharing parameters in several layers. However, in previous approaches, the parameter-sharing techniques have been applied to large deep networks, such as ResNet, that are known to have high redundancy. In this paper, we propose a parameter-sharing method for already parameter-efficient small networks such as ShuffleNetV2. In our approach, small templates are combined with small layer-specific parameters to generate weights. Our experiment results on ImageNet and CIFAR100 datasets show that our approach can reduce the size of parameters by 15%-35% of ShuffleNetV2 while achieving smaller drops in accuracies compared to previous parameter-sharing and pruning approaches. We further show that the proposed approach is efficient in terms of latency and energy consumption on modern embedded devices.

• Journal of Korea Multimedia Society
• /
• v.14 no.2
• /
• pp.318-328
• /
• 2011

This paper suggests a comprehensive method to describe architectural components for supporting Korean Traditional Building design with only small components set in CAD system. Korean traditional buildings can be classified variously based on the their size, usage and structure type(whether ornament part, namely Gongpo, is in there or not). Moreover components can be varied according to the combining rule between them. If all of these components are presented, these tremendous components rather prevent the efficient design of traditional buildings. In order to solve this problem we present object-oriented approach to describe versatile components as one template if they are same in functional aspects. From the template, many similar instances can be derived according to the attribute value. The templates are designed in order to reflect the coupling scheme between components in the relative parameters of the templates. It leads effects of minimizing error which can be occurred frequently in the process of traditional building design.

• 한국HCI학회:학술대회논문집
• /
• 2009.02a
• /
• pp.94-98
• /
• 2009

A video caption recognition system has been developed for broadcasting sport events such as major league baseball. The purpose of the system is to translate the information expressed in English units such as miles per hour (MPH) to the international system of units (SI) such as km/h. The system detects the ball speed displayed in the video and recognizes the numerals. The ball speed is then converted to km/h and displayed by the following character generator (CG) system. Although neural-network based methods are widely used for character and numeral recognition, we use template matching to avoid the training process required before the broadcasting. With the proposed template matching method, the operator can cope with the situation when the caption’s appearance changed without any notification. Templates are configured by the operator with a captured screenshot of the first pitch with ball speed. Templates are updated with following correct recognition results. The accuracy of the recognition module is over 97%, which is still not enough for live broadcasting. When the recognition confidence is low, the system asks the operator for the correct recognition result. The operator chooses the right one using hot keys.

• Journal of the Korean Society for Aeronautical & Space Sciences
• /
• v.48 no.10
• /
• pp.783-789
• /
• 2020

This paper proposes a new method for vision-based navigation using semantically segmented aerial images. Vision-based navigation can reinforce the vulnerability of the GPS/INS integrated navigation system. However, due to the visual and temporal difference between the aerial image and the database image, the existing image matching algorithms have difficulties being applied to aerial navigation problems. For this reason, this paper proposes a suitable matching method for the flight composed of navigational feature extraction through semantic segmentation followed by template matching. The proposed method shows excellent performance in simulation and even flight situations.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.34 no.1C
• /
• pp.81-89
• /
• 2009

We propose a D-ATR UWB (Differential-Average Transmitted Reference Ultra Wide Band) system based on impulse radio. The TR-UWB systems including traditional TR (Transmitted Reference) and ATR (Average TR), exhibit a problem of reduced data rate, since reference signals are additionally transmitted. To tackle this issue, the transmitter of the proposed D-ATR system employs a differential coding like the conventional D-TR system. In addition, the receiver of the proposed system has the structure that can improve signal-to-noise ratio of the reference template used in the correlation process, by recursively averaging the received reference signals like the conventional ATR system. The simulation results in the IEEE 802.15.4a UWB multipath channel models reveal that the proposed D-ATR system achieves much better bit error rate performance as compared to the conventional D- TR system.