• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.3-8
• /
• 2020

문서 검색은 오래 연구되어 온 자연어 처리의 중요한 분야 중 하나이다. 기존의 키워드 기반 검색 알고리즘 중 하나인 BM25는 성능에 명확한 한계가 있고, 딥러닝을 활용한 의미 기반 검색 알고리즘의 경우 문서가 압축되어 벡터로 변환되는 과정에서 정보의 손실이 생기는 문제가 있다. 이에 우리는 BERT Sparse라는 새로운 문서 검색 모델을 제안한다. BERT Sparse는 쿼리에 포함된 키워드를 활용하여 문서를 매칭하지만, 문서를 인코딩할 때는 BERT를 활용하여 쿼리의 문맥과 의미까지 반영할 수 있도록 고안하여, 기존 키워드 기반 검색 알고리즘의 한계를 극복하고자 하였다. BERT Sparse의 검색 속도는 BM25와 같은 키워드 기반 모델과 유사하여 실시간 서비스가 가능한 수준이며, 성능은 Recall@5 기준 93.87%로, BM25 알고리즘 검색 성능 대비 19% 뛰어나다. 최종적으로 BERT Sparse를 MRC 모델과 결합하여 open domain QA환경에서도 F1 score 81.87%를 얻었다.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.19 no.4
• /
• pp.772-780
• /
• 1994

Vector quantization with two-channel conjugate codebook has been researched as an efficient coding technique that can reduce the computational complexity and codebook storage. This paper proposes FSVQ using two-channel conjugate codebook in order to reduce the number of state codebooks. Input vector in the two-channel conjugate FSVQ is coded with state codebook of a seperated state according to each codebook. In addition, LOT is adopted to obtain to obtain a high coding gain and to reduce blocking effect which appears in the block coding. As a result, although FSVQ can achieve higher data compression ratio than general vector quantization, it has a disadvantage of having a very large number of state codebooks. However FSVQ with two-channel conjugate codebooks can employ a significantly reduced number of state codebooks, even though it has a small loss in the PSNR compared with the conventional FSVQ using one codebook. Moreover FSVQ in the LOT domain can reduce blocking effect and high coding gain compared with FSVQ in the spatial domain.

• The Journal of the Korea Contents Association
• /
• v.18 no.12
• /
• pp.35-45
• /
• 2018

Motion Estimation(ME) has been developed to reduce the redundant data in digital video signal. ME is an important part of video encoding system, However, it requires huge computational complexity of the encoder part, and fast motion search methods have been proposed to reduce huge complexity. Multi- view video is obtained by capturing on a three-dimensional scene with many cameras at different positions and its complexity increases in proportion to the number of cameras. In this paper, we proposed an efficient motion method which chooses a search pattern adaptively by using the temporal-spatial correlation of the block and the characteristics of the block. Experiment results show that the computational complexity reduction of the proposed method over TZ search method and FS method can be up to 70~75% and 99% respectively while keeping similar image quality and bit rates.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.21 no.6
• /
• pp.183-190
• /
• 2011

In this paper, we propose a prototype system for abnormal sound detection and identification which detects and recognizes the abnormal situations by means of analyzing audio information coming in real time from CCTV cameras under surveillance environment. The proposed system is composed of two layers: The first layer is an one-class support vector machine, i.e., support vector data description (SVDD) that performs rapid detection of abnormal situations and alerts to the manager. The second layer classifies the detected abnormal sound into predefined class such as 'gun', 'scream', 'siren', 'crash', 'bomb' via a sparse representation classifier (SRC) to cope with emergency situations. The proposed system is designed in a hierarchical manner via a mixture of SVDD and SRC, which has desired characteristics as follows: 1) By fast detecting abnormal sound using SVDD trained with only normal sound, it does not perform the unnecessary classification for normal sound. 2) It ensures a reliable system performance via a SRC that has been successfully applied in the field of face recognition. 3) With the intrinsic incremental learning capability of SRC, it can actively adapt itself to the change of a sound database. The experimental results with the qualitative analysis illustrate the efficiency of the proposed method.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2007.10a
• /
• pp.422-426
• /
• 2007

CP(Counterpropagation) 알고리즘은 Kohonen의 경쟁 네트워크와 Grossberg의 아웃스타(outstar) 구조의 결합으로 이루어진 것으로 패턴 매칭, 패턴 분류, 통계적인 분석 및 데이터 압축 등 활용분야가 다양하고, 다른 신경망 모델에 비해 학습이 매우 빠르다는 장점이 있다. 하지만 CP 알고리즘은 충분한 경쟁층의 수가 설정되지 않아 경쟁층에서 학습이 불안정하고, 여권 코드와 같이 다양한 패턴으로 그성된 경우에는 패턴들을 정확히 분류할 수 없는 단점이 있다. 그리고 CP 알고리즘은 출력층에서 연결강도를 조정할 때, 학습률에 따라 학습 및 인식 성능이 좌우된다. 따라서 본 논문에서는 패턴 인식 성능을 개선하기 위해 다수의 경쟁층을 설정하고, 입력 벡터와 숭자 뉴런의 대표 벡터간의 차이와 숭자 뉴런의 빈도수를 학습률 조정에 반영하여 학습률을 동적으로 조정하여 경쟁층에서 안정적으로 학습되도록 하고, 출력층의 연결강도 조정시 이전 연결 강도 변화량을 반영하는 모멘텀(momentum)학습법을 적용한 개선된 CP 알고리즘을 제안한다. 학습 성능을 확인하기 위해서 실제 여권에서 추출된 개별 코드를 대상으로 실험한 결과, 본 논문에서 개선한 CP 알고리즘이 기존의 CP 알고리즘보다 패턴 분류의 정확성과 인식 성능이 개선된 것을 확인하였다.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.13 no.6
• /
• pp.673-678
• /
• 2003

In this paper, we proposes a method for face recognition using HMM(hidden Markov model) and wavelet coefficients First, input images are compressed by using the multi-resolution analysis based on the discrete wavelet transform. And then, the wavelet coefficients obtained from each subband are used as feature vectors to construct the HMMs. In the recognition stage, we obtained higher recognition rate by summing of each recognition rate of wavelet subband. The usefulness of the proposed method was shown by comparing with conventional VQ and DCT-HMM ones. The experimental results show that the proposed method is more satisfactory than previous ones.

• Journal of Broadcast Engineering
• /
• v.26 no.5
• /
• pp.652-655
• /
• 2021

Among various inter prediction techniques of Versatile Video Coding (VVC), adaptive motion vector resolution (AMVR) technology has been adopted. However, for AMVR, various MVs should be tested per each coding unit, which needs a computation of rate-distortion cost and results in an increase in encoding complexity. Therefore, in order to reduce the encoding complexity of AMVR, it is necessary to effectively find an optimal AMVR mode. In this paper, we propose a lightweight neural network-based AMVR decision algorithm based on more diverse datasets.

• The Journal of the Acoustical Society of Korea
• /
• v.43 no.2
• /
• pp.243-252
• /
• 2024

Among the Foley sound generation models that have recently begun to be studied, a sound generation technique using the Vector Quantized-Variational AutoEncoder (VQ-VAE) structure and generation model such as Pixelsnail are one of the important research subjects. On the other hand, in the field of deep learning-based acoustic signal compression, residual vector quantization technology is reported to be more suitable than the conventional VQ-VAE structure. Therefore, in this paper, we aim to study whether residual vector quantization technology can be effectively applied to the Foley sound generation. In order to tackle the problem, this paper applies the residual vector quantization technique to the conventional VQ-VAE-based Foley sound generation model, and in particular, derives a model that is compatible with the existing models such as Pixelsnail and does not increase computational resource consumption. In order to evaluate the model, an experiment was conducted using DCASE2023 Task7 data. The results show that the proposed model enhances about 0.3 of the Fréchet audio distance. Unfortunately, the performance enhancement was limited, which is believed to be due to the decrease in the resolution of time-frequency domains in order to do not increase consumption of the computational resources.

• Journal of Korea Multimedia Society
• /
• v.10 no.1
• /
• pp.58-65
• /
• 2007

This dissertation proposes digital video protection algorithm lot moving image based on MPEG. Shuffling-based encryption algorithms using a fixed random shuffling table are quite simple and effective but vulnerable to the chosen plaintext attack. To overcome this problem, it is necessary to change the key used for generation of the shuffling table. However, this may pose a significant burden on the security key management system. A better approach is to generate the shuffling table based on the local feature of an image. In order to withstand the chosen plaintext attack, at first, we propose a interleaving algorithm that is adaptive to the local feature of an image. Secondly, using the multiple shuffling method which is combined interleaving with existing random shuffling method, we encrypted the DPCM processed 8*8 blocks. Experimental results showed that the proposed algorithm needs only 10% time of SEED encryption algorithm and moreover there is no overhead bit. In video sequence encryption, multiple random shuffling algorithms are used to encrypt the DC and AC coefficients of intra frame, and motion vector encryption and macroblock shuffling are used to encrypt the intra-coded macroblock in predicted frame.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.10 no.7
• /
• pp.1214-1221
• /
• 2006

Header compression scheme is suggested as a solution to reduce the inefficient overhead of general packet stream data. Especially, it is shown that there are more overhead rate for real-time media stream links such as voice because of its short payload size, and it is possible to get higher bandwidth efficiency using the header compression scheme. There are two kinds of error recovery in header compression such as Periodic Header Refresh(PHR) and Header Request(HR) schemes. In this paper, we analyze the performance of these two compression recovery schemes, and some results such as the overhead rate, bandwidth gain and bandwidth efficiency(BE) are presented.