• 한국음향학회지
• /
• 제40권5호
• /
• pp.488-495
• /
• 2021

본 논문에서는 콘포머 기반 한국어 음성인식 시스템을 제안한다. 콘포머는 트랜스포머 모델에 콘볼루션신경망(Convolution Neural Network, CNN) 기능을 보강한 구조이며 광역 정보를 잘 표현할 수 있는 트랜스포머와 지역 정보를 잘 표현할 수 있는 CNN을 결합한 신경망이다. 음성인식 기본 시스템으로 트랜스포모에 기반한 음성인식시스템을 개발하였으며 언어모델로는 Long Short-Term Memory(LSTM)을 사용하였다. 콘포머 기반 음성인식시스템은 트랜스포머 대신에 콘포머를 사용하였고 언어모델로는 트랜스포머를 이용하였다. 성능 평가를 위해 AI-hub에 있는 Electronics and Telecommunications Research Institute(ETRI) 음성코퍼스를 활용하였으며 트랜스포머 기반 음성인식 시스템은 오인식률이 11.8 %이 되었으며 콘포머 기반 음성인식시스템은 오인식률이 5.7 %가 되었다. AI-hub에 있는 다른 영역의 NHN다이퀘스트 음성 코퍼스를 추가해도 유사한 성능이 유지가 되어 제안된 콘포머 음성인식시스템의 유효성을 입증하였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제13권1호
• /
• pp.371-384
• /
• 2019

Motion perception has been tremendously improved in neuroscience and computer vision. The baseline motion perception model is mediated by the dorsal visual pathway involving the cortex areas the primary visual cortex (V1) and the middle temporal (V5 or MT) visual area. However, few works have been done on the extension of neural models to improve the efficacy and robustness of motion perception of real sequences. To overcome shortcomings in situations, such as varying illumination and large displacement, an adaptive V1-MT motion perception (Ad-V1MTMP) algorithm enriched to deal with real sequences is proposed and analyzed. First, the total variation semi-norm model based on Gabor functions (TV-Gabor) for structure-texture decomposition is performed to manage the illumination and color changes. And then, we study the impact of image local context, which is processed in extra-striate visual areas II (V2), on spatial motion integration by MT neurons, and propose a V1-V2 method to extract the image contrast information at a given location. Furthermore, we take feedback inputs from V2 into account during the polling stage. To use the algorithm on natural scenes, finally, multi-scale approach has been used to handle the frequency range, and adaptive pyramidal decomposition and decomposed spatio-temporal filters have been used to diminish computational cost. Theoretical analysis and experimental results suggest the new Ad-V1MTMP algorithm which mimics human primary motion pathway has universal, effective and robust performance.

• 한국음향학회지
• /
• 제38권1호
• /
• pp.51-55
• /
• 2019

본 논문에서는 잡음 환경에 강인한 음성 인식 성능을 위해 특징 보상 이득을 이용한 음성 향상 기법을 제안한다. 본 논문에서는 변분모델 생성 기법을 채용한 병렬 결합된 가우스 혼합 모델(Parallel Combined Gaussian Mixture Model, PCGMM) 기반의 특징 보상 기법으로부터 계산할 수 있는 특징 보상 이득을 이용하는 음성 향상 기술을 제안한다. 불일치 환경 음성 인식 시스템 적용 환경에서 본 논문에서 제안하는 기법이 실험 결과에서 기존의 전처리 기법 및 이전 연구에서 제안된 특징 보상 기반의 음성 향상 기법에 비해 다양한 잡음 및 SNR(Signal to Noise Ratio) 조건에서 월등한 인식 성능을 나타내는 것을 확인한다. 또한 잡음 모델 선택 기법을 적용함으로써 음성 인식 성능을 유사한 수준으로 유지하면서 계산량을 대폭적으로 감축할 수 있다.

• 한국정보기술학회논문지
• /
• 제16권11호
• /
• pp.97-103
• /
• 2018

본 논문에서는 시변 최적 FIR 필터를 이용한 전력 신호의 고조파 검출 방법을 제안한다. 잡음이 고려된 전력 신호에 대해 고조파의 진폭과 위상각의 변화량이 확률적 정보로 고려된 시변 상태 방정식 모델에 대해 시변 최적 FIR 필터를 설계하여 고조파 성분을 검출한다. 제안한 검출 방법에 사용된 시변 FIR 필터는 유한 구간의 정보만을 사용하고 어떠한 초기 조건도 사용하지 않도록 설계되어 칼만 필터 기반의 검출 방법의 오차 누적에 따른 검출 성능 저하나 발산 문제를 해결할 수 있다. 또한 FIR 구조의 필터 사용을 통해 칼만 필터 대비 불확실성에 대해 보다 강인한 검출이 가능하다. 시변 최적 FIR 필터의 사용을 통해 시불변 최적 FIR 필터 기반 고조파 검출 방법 대비 보다 일반적인 해를 제공한다. 제안하는 검출 방법의 우수성을 검증하기 위해 시변 칼만 필터 및 적응 칼만 필터 기반 고조파 검출 방법과의 비교 시뮬레이션을 수행한다.

• 한국산학기술학회논문지
• /
• 제20권9호
• /
• pp.49-55
• /
• 2019

본 논문에서는 효과적인 마커기반의 증강현실 구현을 위하여 영상 내 객체의 분포에 대한 분석과 반복 패턴을 포함하는 영상의 분류를 통한 마커영상의 평가 방법을 제안한다. 객체의 분포는 영상의 부분적 가림 현상에 따라 객체추적성능에 영향을 미치기 때문에 특징점 좌표의 분산을 이용하여 가림 현상에 취약한 마커영상을 구분할 수 있도록 하였고, 일반 영상과 반복 패턴을 포함하는 영상의 특징점 기술자 벡터의 분포가 현저하게 다르다는 사실에 기반하여 객체의 인식 및 추적에 적합한 영상을 구분할 수 있는 방법을 제안한다. 다양한 실험 결과 제안하는 마커 평가 방법이 가림 현상에 취약한 영상 및 반복 패턴 영상을 구분하는데 우수한 성능을 보이는 것을 확인하였다. 또한 마커영상에 대한 객체 추적 등의 안정성 측면에서 SURF보다 SIFT 기법이 우수한 성능을 보임을 확인할 수 있었다. 이러한 결과를 이용하여 다양한 종류의 마커영상에 대한 적합성 정보를 사용자에게 제공함으로써 효과적인 증강현실 시스템을 구현할 수 있을 것으로 판단된다.

• Corrosion Science and Technology
• /
• 제18권2호
• /
• pp.61-71
• /
• 2019

Flow-Accelerated Corrosion (FAC) is a phenomenon in which a protective coating on a metal surface is dissolved by a flow of fluid in a metal pipe, leading to continuous wall-thinning. Recently, many countries have developed computer codes to manage FAC in power plants, and the FAC prediction model in these computer codes plays an important role in predictive performance. Herein, the FAC prediction model was developed by applying a machine learning method and the conventional nonlinear regression method. The random forest, a widely used machine learning technique in predictive modeling led to easy calculation of FAC tendency for five input variables: flow rate, temperature, pH, Cr content, and dissolved oxygen concentration. However, the model showed significant errors in some input conditions, and it was difficult to obtain proper regression results without using additional data points. In contrast, nonlinear regression analysis predicted robust estimation even with relatively insufficient data by assuming an empirical equation and the model showed better predictive power when the interaction between DO and pH was considered. The comparative analysis of this study is believed to provide important insights for developing a more sophisticated FAC prediction model.

• Journal of Microbiology and Biotechnology
• /
• 제31권8호
• /
• pp.1154-1162
• /
• 2021

The transcriptional capacities of target genes are strongly influenced by promoters, whereas few studies have focused on the development of robust, high-performance and cross-species promoters for wide application in different bacteria. In this work, four novel promoters (P_k.rtufB, P_k.r1, P_k.r2, and P_k.r3) were predicted from Ketogulonicigenium robustum and their inconsistency in the -10 and -35 region nucleotide sequences indicated they were different promoters. Their activities were evaluated by using green fluorescent protein (gfp) as a reporter in different species of bacteria, including K. vulgare SPU B805, Pseudomonas putida KT2440, Paracoccus denitrificans PD1222, Bacillus licheniformis and Raoultella ornithinolytica, due to their importance in metabolic engineering. Our results showed that the four promoters had different activities, with P_k.r1 showing the strongest activity in almost all of the experimental bacteria. By comparison with the commonly used promoters of E. coli (tufB, lac, lacUV5), K. vulgare (Psdh, Psndh) and P. putida KT2440 (JE111411), the four promoters showed significant differences due to only 12.62% nucleotide similarities, and relatively higher ability in regulating target gene expression. Further validation experiments confirmed their ability in initiating the target minCD cassette because of the shape changes under the promoter regulation. The overexpression of sorbose dehydrogenase and cytochrome c551 by P_k.r1 and P_k.r2 resulted in a 22.75% enhancement of 2-KGA yield, indicating their potential for practical application in metabolic engineering. This study demonstrates an example of applying bioinformatics to find new biological components for gene operation and provides four novel promoters with broad suitability, which enriches the usable range of promoters to realize accurate regulation in different genetic backgrounds.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권3호
• /
• pp.837-852
• /
• 2021

Recurrent neural network (RNN) architectures have been used for language modeling (LM) tasks that require learning long-range word or character sequences. However, the RNN architecture is still suffered from unstable gradients on long-range sequences. To address the issue of long-range sequences, an attention mechanism has been used, showing state-of-the-art (SOTA) performance in all LM tasks. A differentiable neural computer (DNC) is a deep learning architecture using an attention mechanism. The DNC architecture is a neural network augmented with a content-addressable external memory. However, in the write operation, some information unrelated to the input word remains in memory. Moreover, DNCs have been found to perform poorly with low numbers of weight parameters. Therefore, we propose a robust memory deallocation method using a limited retention vector. The limited retention vector determines whether the network increases or decreases its usage of information in external memory according to a threshold. We experimentally evaluate the robustness of a DNC implementing the proposed approach according to the size of the controller and external memory on the enwik8 LM task. When we decreased the number of weight parameters by 32.47%, the proposed DNC showed a low bits-per-character (BPC) degradation of 4.30%, demonstrating the effectiveness of our approach in language modeling tasks.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제15권3호
• /
• pp.1078-1099
• /
• 2021

Traditional image steganography hides secret information by embedding, which inevitably leaves modification traces and is easy to be detected by steganography analysis tools. Since coverless steganography can effectively resist steganalysis, it has become a hotspot in information hiding research recently. Most coverless image steganography (CIS) methods are based on mapping rules, which not only exposes the vulnerability to geometric attacks, but also are less secure due to the revelation of mapping rules. To address the above issues, we introduced camouflage images for steganography instead of directly sending stego-image, which further improves the security performance and information hiding ability of steganography scheme. In particular, based on the different sub-features of stego-image and potential camouflage images, we try to find a larger similarity between them so as to achieve the reversible steganography. Specifically, based on the existing CIS mapping algorithm, we first can establish the correlation between stego-image and secret information and then transmit the camouflage images, which are obtained by reversible sub-feature retrieval algorithm. The received camouflage image can be used to reverse retrieve the stego-image in a public image database. Finally, we can use the same mapping rules to restore secret information. Extensive experimental results demonstrate the better robustness and security of the proposed approach in comparison to state-of-art CIS methods, especially in the robustness of geometric attacks.

• Computers and Concrete
• /
• 제27권4호
• /
• pp.319-332
• /
• 2021

The performance of gene expression programming (GEP) in predicting the compressive strength of bacteria-incorporated geopolymer concrete (GPC) was examined in this study. Ground-granulated blast-furnace slag (GGBS), new bacterial strains, fly ash (FA), silica fume (SF), metakaolin (MK), and manufactured sand were used as ingredients in the concrete mixture. For the geopolymer preparation, an 8 M sodium hydroxide (NaOH) solution was used, and the ambient curing temperature (28℃) was maintained for all mixtures. The ratio of sodium silicate (Na2SiO3) to NaOH was 2.33, and the ratio of alkaline liquid to binder was 0.35. Based on experimental data collected from the literature, an evolutionary-based algorithm (GEP) was proposed to develop new predictive models for estimating the compressive strength of GPC containing bacteria. Data were classified into training and testing sets to obtain a closed-form solution using GEP. Independent variables for the model were the constituent materials of GPC, such as FA, MK, SF, and Bacillus bacteria. A total of six GEP formulations were developed for predicting the compressive strength of bacteria-incorporated GPC obtained at 1, 3, 7, 28, 56, and 90 days of curing. 80% and 20% of the data were used for training and testing the models, respectively. R2 values in the range of 0.9747 and 0.9950 (including train and test dataset) were obtained for the concrete samples, which showed that GEP can be used to predict the compressive strength of GPC containing bacteria with minimal error. Moreover, the GEP models were in good agreement with the experimental datasets and were robust and reliable. The models developed could serve as a tool for concrete constructors using geopolymers within the framework of this research.