• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.10a
• /
• pp.763-766
• /
• 2018

표정은 사람들 사이에서 감정을 표현하는 강력한 비언어적 수단이다. 표정 인식은 기계학습에서 아주 중요한 분야 중에 하나이다. 표정 인식에 사용되는 기계학습 모델들은 사람 수준의 성능을 보여준다. 하지만 좋은 성능에도 불구하고, 기계학습 모델들은 표정 인식 결과에 대한 근거나 설명을 제공해주지 못한다. 이 연구는 표정 인식의 근거로서 Facial Action Coding Unit(AUs)을 사용하기 위해서 CK+ Dataset을 사용하여 표정 인식을 학습한 Convolutional Neural Network(CNN) 모델이 추출한 특징들을 t-distributed stochastic neighbor embedding(t-SNE)을 사용하여 시각화한 뒤, 인식된 표정과 AUs 사이의 분포의 연관성을 확인하는 연구이다.

• The KIPS Transactions:PartB
• /
• v.10B no.7
• /
• pp.777-784
• /
• 2003

In this Paper. we propose a recognition method of machine printed characters based on the combination of recognition units using multiple neural networks. In our recognition method, the input character is classified into one of 7 character types among which the first 6 types are for Hangul character and the last type is for non-Hangul characters. Hangul characters are recognized by several MLP (multilayer perceptron) neural networks through two stages. In the first stage, we divide Hangul character image into two or three recognition units (HRU : Hangul recognition unit) according to the combination fashion of graphemes. Each recognition unit composed of one or two graphemes is recognized by an MLP neural network with an input feature vector of pixel direction angles. In the second stage, the recognition aspect features of the HRU MLP recognizers in the first stage are extracted and forwarded to a subsequent MLP by which final recognition result is obtained. For the recognition of non-Hangul characters, a single MLP is employed. The recognition experiments had been performed on the character image database collected from 50,000 real letter envelope images. The experimental results have demonstrated the superiority of the proposed method.

• KIPS Transactions on Software and Data Engineering
• /
• v.7 no.7
• /
• pp.259-266
• /
• 2018

In this paper, we propose a method to classify a document using a Recurrent Neural Network by extracting features considering word sense and contexts. Word2vec method is adopted to include the order and meaning of the words expressing the word in the document as a vector. Doc2vec is applied for considering the context to extract the feature of the document. RNN classifier, which includes the output of the previous node as the input of the next node, is used as the document classification method. RNN classifier presents good performance for document classification because it is suitable for sequence data among neural network classifiers. We applied GRU (Gated Recurrent Unit) model which solves the vanishing gradient problem of RNN. It also reduces computation speed. We used one Hangul document set and two English document sets for the experiments and GRU based document classifier improves performance by about 3.5% compared to CNN based document classifier.

• KIISE Transactions on Computing Practices
• /
• v.23 no.12
• /
• pp.698-702
• /
• 2017

Since the introduction of Deep Neural Networks to the Natural Language Processing field, two major approaches have been considered for modeling text. One method involved learning embeddings, i.e. the distributed representations containing abstract semantics of words or sentences, with the textual context. The other strategy consisted of composing the embeddings trained by the above to get embeddings of longer texts. However, most studies of the composition methods just adopt word embeddings without consideration of the optimal embedding unit and the optimal method of composition. In this paper, we conducted experiments to analyze the optimal embedding unit and the optimal composition method for modeling longer texts, such as documents. In addition, we suggest a new discourse-based composition to overcome the limitation of the sequential composition method on composing sentence embeddings.

• Journal of Information Processing Systems
• /
• v.9 no.1
• /
• pp.117-140
• /
• 2013

Hindi is the most widely spoken language in India, with more than 300 million speakers. As there is no separation between the characters of texts written in Hindi as there is in English, the Optical Character Recognition (OCR) systems developed for the Hindi language carry a very poor recognition rate. In this paper we propose an OCR for printed Hindi text in Devanagari script, using Artificial Neural Network (ANN), which improves its efficiency. One of the major reasons for the poor recognition rate is error in character segmentation. The presence of touching characters in the scanned documents further complicates the segmentation process, creating a major problem when designing an effective character segmentation technique. Preprocessing, character segmentation, feature extraction, and finally, classification and recognition are the major steps which are followed by a general OCR. The preprocessing tasks considered in the paper are conversion of gray scaled images to binary images, image rectification, and segmentation of the document's textual contents into paragraphs, lines, words, and then at the level of basic symbols. The basic symbols, obtained as the fundamental unit from the segmentation process, are recognized by the neural classifier. In this work, three feature extraction techniques-: histogram of projection based on mean distance, histogram of projection based on pixel value, and vertical zero crossing, have been used to improve the rate of recognition. These feature extraction techniques are powerful enough to extract features of even distorted characters/symbols. For development of the neural classifier, a back-propagation neural network with two hidden layers is used. The classifier is trained and tested for printed Hindi texts. A performance of approximately 90% correct recognition rate is achieved.

• Journal of Information Processing Systems
• /
• v.19 no.6
• /
• pp.730-744
• /
• 2023

The topic of this paper is the recognition of human activities using egocentric vision, particularly captured by body-worn cameras, which could be helpful for video surveillance, automatic search and video indexing. This being the case, it could also be helpful in assistance to elderly and frail persons for revolutionizing and improving their lives. The process throws up the task of human activities recognition remaining problematic, because of the important variations, where it is realized through the use of an external device, similar to a robot, as a personal assistant. The inferred information is used both online to assist the person, and offline to support the personal assistant. With our proposed method being robust against the various factors of variability problem in action executions, the major purpose of this paper is to perform an efficient and simple recognition method from egocentric camera data only using convolutional neural network and deep learning. In terms of accuracy improvement, simulation results outperform the current state of the art by a significant margin of 61% when using egocentric camera data only, more than 44% when using egocentric camera and several stationary cameras data and more than 12% when using both inertial measurement unit (IMU) and egocentric camera data.

• The Journal of the Acoustical Society of Korea
• /
• v.17 no.4E
• /
• pp.28-34
• /
• 1998

In most switching system, the processing unit is designed to work efficiently even at relatively high loads, but when the offered traffic exceeds a particular level, the rate of completed calls can fall drastically. A single call handled by the switching system consists of a sequence of events or messages that has to be processed by the control unit. The control unit is not only incapable of handling all of the offered calls, but also its call handling capability can drop as the offered load increases. The real time available for call processing is a critical resource that requires careful management. Therefore, the overloading of this resource must be detected by a subscriber in the from of a dial tone delay or an uncompleted call which is either blocked or mishandled. The subscriber may respond by either dialing prematurely or by re-attempting a call. This action can further escalate the processors load, which is spent for uncompleted calls. Unless a proper control is used, the switching system can finally break down. In this paper, we paper, we propose a fuzzy overload detection and control method for switching systems, which can by generating fuzzy rules via fuzzy aggregation networks. Simulation results involving a switching system is given.

• Electronics and Telecommunications Trends
• /
• v.32 no.6
• /
• pp.57-65
• /
• 2017

Neural network based AI computing is a promising technology that reflects the recognition and decision operation of human beings. Early AI computing processors were composed of GPUs and CPUs; however, the dramatic increment of a floating point operation requires an energy efficient AI processor with a highly parallelized architecture. In this paper, we analyze the trends in processor architectures for AI computing. Some architectures are still composed using GPUs. However, they reduce the size of each processing unit by allowing a half precision operation, and raise the processing unit density. Other architectures concentrate on matrix multiplication, and require the construction of dedicated hardware for a fast vector operation. Finally, we propose our own inAB processor architecture and introduce domestic cutting-edge processor design capabilities.

• IEIE Transactions on Smart Processing and Computing
• /
• v.6 no.3
• /
• pp.193-199
• /
• 2017

Instead of using only spatial features from a single frame for person re-identification, a combination of spatial and temporal factors boosts the performance of the system. A recurrent neural network (RNN) shows its effectiveness in generating highly discriminative sequence-level human representations. In this work, we implement RNN, three Long Short Term Memory (LSTM) network variants, and Gated Recurrent Unit (GRU) on Caffe deep learning framework, and we then conduct experiments to compare performance in terms of size and accuracy for person re-identification. We propose using GRU for the optimized choice as the experimental results show that the GRU achieves the highest accuracy despite having fewer parameters than the others.

• The Transactions of the Korean Institute of Electrical Engineers P
• /
• v.58 no.4
• /
• pp.511-515
• /
• 2009

For detecting of the faulty insulator, Correlation Propagation Neural Networks(CPNN) has been proposed. Faulty insulator is reduced the rate of insulation extremely, and taken the results dirty and injured. It is necessary to detect the faulty insulator and exchange the new one. And thus, we have designed the CPNN to be detected that insulators by the real time computation method through the inter-node diffusion. In the network, a node corresponds to a state in the quantized input space. Each node is composed of a processing unit and fixed weights from its neighbor nodes as well as its input terminal. Information propagates among neighbor nodes laterally and inter-node interpolation is achieved. 1-D CPNN hardware has been implemented with general purpose. Experiments with static and dynamic signals have been done upon the CPNN hardware. Through the results of simulation experiments, we define the ability of real-time detecting the faulty insulators.