• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.13 no.11
• /
• pp.5427-5445
• /
• 2019

While deep neural networks have achieved remarkable performance in representation learning, a huge amount of labeled training data are usually required by supervised deep models such as convolutional neural networks. In this paper, we propose a new representation learning method, namely generative adversarial networks (GAN) based bagging deep convolutional autoencoders (GAN-BDCAE), which can map data to diverse hierarchical representations in an unsupervised fashion. To boost the size of training data, to train deep model and to aggregate diverse learning machines are the three principal avenues towards increasing the capabilities of representation learning of neural networks. We focus on combining those three techniques. To this aim, we adopt GAN for realistic unlabeled sample generation and bagging deep convolutional autoencoders (BDCAE) for robust feature learning. The proposed method improves the discriminative ability of learned feature embedding for solving subsequent pattern recognition problems. We evaluate our approach on three standard benchmarks and demonstrate the superiority of the proposed method compared to traditional unsupervised learning methods.

• The Journal of Korea Robotics Society
• /
• v.15 no.1
• /
• pp.55-61
• /
• 2020

Getting on and off an elevator is one of the most important parts for multi-floor navigation of a mobile robot. In this study, we proposed the method for the pose recognition of elevator doors, safe path planning, and motion estimation of a robot using RGB-D sensors in order to safely get on and off the elevator. The accurate pose of the elevator doors is recognized using a particle filter algorithm. After the elevator door is open, the robot builds an occupancy grid map including the internal environments of the elevator to generate a safe path. The safe path prevents collision with obstacles in the elevator. While the robot gets on and off the elevator, the robot uses the optical flow algorithm of the floor image to detect the state that the robot cannot move due to an elevator door sill. The experimental results in various experiments show that the proposed method enables the robot to get on and off the elevator safely.

• The Journal of Korea Robotics Society
• /
• v.7 no.2
• /
• pp.92-100
• /
• 2012

A map of complex environment can be generated using a robot carrying sensors. However, representation of environments directly using the integration of sensor data tells only spatial existence. In order to execute high-level applications, robots need semantic knowledge of the environments. This research investigates the design of a system for recognizing objects in 3D point clouds of urban environments. The proposed system is decomposed into five steps: sequential LIDAR scan, point classification, ground detection and elimination, segmentation, and object classification. This method could classify the various objects in urban environment, such as cars, trees, buildings, posts, etc. The simple methods minimizing time-consuming process are developed to guarantee real-time performance and to perform data classification on-the-fly as data is being acquired. To evaluate performance of the proposed methods, computation time and recognition rate are analyzed. Experimental results demonstrate that the proposed algorithm has efficiency in fast understanding the semantic knowledge of a dynamic urban environment.

• The Journal of the Acoustical Society of Korea
• /
• v.14 no.1
• /
• pp.13-21
• /
• 1995

In this paper, we improve an image classifier algorithm based on neural network learning. It consists of two steps. The first is input pattern generation and the second, the global neural network implementation using an improved back-propagation algorithm. The feature vector for pattern recognition consists of the codebook data obtained from self-organization feature map learning. It decreases the input neuron number as well as the computational cost. The global neural network algorithm which is used in classifier inserts a control part and an address memory part to the back-propagation algorithm to control weights and unit-offsets. The simulation results show that it does not fall into the local minima and can implement easily the large-scale neural network. And it decreases largely the learning time.

• Journal of Auto-vehicle Safety Association
• /
• v.7 no.4
• /
• pp.26-30
• /
• 2015

Exceeding the speed limit during vehicle driving is a key factor in the severity of lots of road accidents, and SLIF(Speed Limit Information Function) application is in the initial phase in Korea. SLIF helps the drivers to observe a speed limit when they are driving by providing alert and informing the current limit speed information based on external data using camera and/or digital map, for that reason, environmental conditions could be causes of SLIF malfunctions. In this study, design adequacy analysis of SLIF in respect of false recognition as the Korea traffic environment has been performed. As tentative results, road conditions and structure of speed limit sign as well as system performance often caused misrecognition.

• International Journal of Internet, Broadcasting and Communication
• /
• v.9 no.4
• /
• pp.1-7
• /
• 2017

Clipart is artificial visual contents that are created using various tools such as Illustrator to highlight some information. Here, the style of the clipart plays a critical role in determining how it looks. However, previous studies on clipart are focused only on the object recognition [16], segmentation, and retrieval of clipart images using hand-craft image features. Recently, some clipart classification researches based on the style similarity using CNN have been proposed, however, they have used different CNN-models and experimented with different benchmark dataset so that it is very hard to compare their performances. This paper presents an experimental analysis of the clipart classification based on the style similarity with two well-known CNN-models (Inception Resnet V2 [13] and VGG-16 [14] and transfers learning with the same benchmark dataset (Microsoft Style Dataset 3.6K). From this experiment, we find out that the accuracy of Inception Resnet V2 is better than VGG for clipart style classification because of its deep nature and convolution map with various sizes in parallel. We also find out that the end-to-end training can improve the accuracy more than 20% in both CNN models.

• Journal of Institute of Control, Robotics and Systems
• /
• v.15 no.12
• /
• pp.1232-1239
• /
• 2009

This paper proposes an exploration strategy to efficiently find a specific place in large unknown environments with wall-following based path planning. Many exploration methods proposed so far showed good performance but they focused only on efficient planning for modeling unknown environments. Therefore, to successfully accomplish the room finding task, two additional requirements should be considered. First, suitable path-planning is needed to recognize the room number. Most conventional exploration schemes used the gradient method to extract the optimal path. In these schemes, the paths are extracted in the middle of the free space which is usually far from the wall. If the robot follows such a path, it is not likely to recognize the room number written on the wall because room numbers are usually too small to be recognized by camera image from a distance. Second, the behavior which re-explores the explored area is needed. Even though the robot completes exploration, it is possible that some rooms are not registered in the constructed map for some reasons such as poor recognition performance, occlusion by a human and so on. With this scheme, the robot does not have to visit and model the whole environment. This proposed method is very simple but it guarantees that the robot can find a specific room in most cases. The proposed exploration strategy was verified by various experiments.

• Transactions of the Korean Society of Mechanical Engineers A
• /
• v.21 no.10
• /
• pp.1551-1560
• /
• 1997

This study is devoted to the detection of the 3-dimensional point obstacles on the plane by using accumulated scan line images. The proposed accumulating only one scan line allow to process image at real time. And the change of motion of the feature in image is small because of the short time between image frames, so it does not take much time to track features. To obtain recursive optimal obstacles position and robot motion along to the motion of camera, Kalman filter algorithm is used. After using Kalman filter in case of the fixed environment, 3-dimensional obstacles point map is obtained. The position and motion of moving obstacles can also be obtained by pre-segmentation. Finally, to solve the stereo ambiguity problem from multiple matches, the camera motion is actively used to discard mis-matched features. To get relative distance of obstacles from camera, parallel stereo camera setup is used. In order to evaluate the proposed algorithm, experiments are carried out by a small test vehicle.

• Annual Conference on Human and Language Technology
• /
• 1996.10a
• /
• pp.145-152
• /
• 1996

연속음성인식 시스템을 개발하기 위해서는 언어가 갖는 문법적 제약을 이용한 언어모델이 요구된다. 문법적 규칙을 이용한 언어모델은 전문가가 일일이 문법 규칙을 만들어 주어야 하는 단점이 있다. 통계적 언어 모델에서는 문법적인 정보를 수작업으로 만들어 주지 않는 대신 그러한 모든 정보를 학습을 통해서 훈련해야 하기 때문에 이를 위해 요구되는 학습 데이터도 엄청나게 증가한다. 따라서 적은 양의 데이터로도 이와 유사한 효과를 보일 수 있는 것이 클래스에 의거한 언어 모델이다. 또 이 모델은 음성 인식과 연계시에 탐색 공간을 줄여 주기 때문에 실시간 시스템 구현에 매우 유용한 모델이다. 여기서는 자동으로 클래스를 찾아주는 알고리즘을 호텔예약시스템의 corpus에 적용, 분석해 보았다. Corpus 자체가 문법규칙이 뚜렷한 특성을 갖고 있기 때문에 heuristic하게 클래스를 준 것과 유사한 결과를 보였지만 corpus 크기가 커질 경우에는 매우 유용할 것이며, initial map을 heuristic하게 주고 그 알고리즘을 적용한 결과 약간의 성능향상을 볼 수 있었다. 끝으로 음성인식시스템과 접합해 본 결과 유사한 결과를 얻었으며 언어모델에도 음향학적 특성을 반영할 수 있는 연구가 요구됨을 알 수 있었다.

• Proceedings of the KIEE Conference
• /
• 2005.11a
• /
• pp.9-11
• /
• 2005

본 연구는 GIS 진단신뢰성 향상기술 개발을 목적으로, 16개의 인위적 결함을 이용하여 부분방전 신호를 발생시키고 검출하여 그 패턴인식 확률을 높이기 위하여 신경망에 Genetic Algorithm (GA) 을 적용하였다. 이를 위하여 다음과 같은 5가지 서로 다른 신경망 모델을 선택하였다: Back Propagation (BP), Jordan-Elman Network (JEN), Principal Component Analysis (PCA), Self-Organizing Feature Map (SOFM) 및 Support Vector Machine (SVM). 이와 같이 선택된 모델에 동일한 데이터를 학습 시키고 패턴인식 확률을 비교 및 분석하였다. 실험 결과에 의하면, BP의 인식률이 가장 높고 다음으로 JEN의 인식률이 높이 나타났으며, 후자의 경우 모든 결함에 대하여 정확한 패턴분류를 한 반면에 전자의 경우 1.8% 의 분류 오차가 발생하였다. 따라서 인식률이 높은 신경망이 더 정확한 패턴분류를 보장하지 못한다는 실험적 결과를 고려 할 때, 인식률이 높은 두 개의 모델을 선정하여 각각의 출력에 일정한 가중치를 주고 합산하여 새로운 출력을 얻는 방법을 제안한다.