Earticle

본 논문에서는 기존의 인공신경망을 이용한 자동작곡 방법에서 발생한 세 가지 문제점을 개선하는 새로운 방법을 제안한다. 첫 번째 문제는 인공신경망이 출력한 곡의 박자를 음악이론에 맞도록 후처리 하는 것에서 모든 경우를 처리하지 못하여 발생한 문제이다. 두 번째 문제는 음표를 학습하는 인공신경망에 음표와 구분되는 큰 값으로 쉼표를 같이 학습하다보니 음표공간이 왜곡되어 발생하는 문제이다. 마지막 문제는 새로운 곡 작곡 시 사용자가 작곡해서 넣어준 초기 멜로디와 박자가 인공신경망이 출력하는 나머지 멜로디와 박자와 어울리지 못하여 발생하는 문제이다. 본 논문에서는 이러한 문제를 해결하기 위하여 개선된 박자 후처리 알고리즘과 초기 멜로디 처리 방법을 제안하였으며 쉼표용 인공신경망을 새로이 도입하였다. 실험결과 새로 제안한 방법이 기존의 방법에서 발생한 세 가지 문제점을 모두 해결하는 것으로 판명되었다.

This paper proposes a new method to improve the three problems of existing automatic composition method using artificial neural networks. The first problem is that the existing beat post-processing to fit into music theories could not handle all the cases of occurring. The second one is that the pitch space generated by artificial neural networks is distorted because the rest is trained with the pitch on the same neural network with large values. The last problem is caused by the difference between the initial melody and beats given by user and those generated by an artificial neural network in the process of new composition. In order to treat these problems, we propose an enhanced post-processing of beats, initial melody processing, and adoption of artificial neural network for rest. It was found from experiments that the proposed methods totally resolved the three problems.

This paper presents an adaptive interpolation algorithm for effective picture format conversion, namely deinterlacing based on total edge directions estimation. The deinterlacing algorithms are generally made up of two parts, such as the estimation of edge directions and the interpolation of missing pixels along with the estimated direction. The proposed algorithm adopts and enhances the block-based search concept for estimation of edge directions, which is introduced in the direction-oriented interpolation (DOI) algorithm. In this paper the edge directions are estimated with the vertical angles as well as the horizontal angles in the DOI and then the missing pixels are interpolated by using median operation. Experimental results show that the proposed algorithm can estimate edge directions and interpolate missing pixels more accurately compared with the conventional deinterlacing algorithms in terms of the objective performance measured in PSNR and the subjective visual quality.

이 연구는 제한된 용량과 기능의 마이크로컴퓨터를 전제로하여 고급의 postprocessing을 효율적으로 달성할 수 있는 algorithm을 개발하고자 의도되었다. 응력도 및 내력도의 도화, 변형의 animation등과 관련된 기존의 postprocessing방법을 개선하엿으며, 이를 바탕으로 새로운 유한요소해석시스템을 개발하였다. 특히 3차원 구조물의 역학적 거동을 가시화하는 문제를 중점적으로 다루었다.

This study was intended to provide efficient algorithms for high quality postprocessing on the basis of microcomputers with limited capacity and functionality. Improved methods of postprocessing including stress contouring, internal force diagraming, and displacement animation, were proposed and implemented into a new finite element system. Visualization of three-dimensional structural behaviors was treated with special emphasis.

This study was intended to provide efficient algorithms for high quality postprocess on the basis of microcomputers with limited capacity and functionality. Improved methods of postprocessing including stress contouring, internal force diagraming, and displacement animation, were proposed and implemented into a new finite element system. Visualization of three-dimensional structural behaviors was treated with special emphasis.

In this paper we present a postprocessing scheme for the Raviart-Thomas mixed finite element approximation of the second order elliptic eigenvalue problem. This scheme is carried out by solving a primal source problem on a higher order space, and thereby can improve the convergence rate of the eigenfunction and eigenvalue approximations. It is also used to compute a posteriori error estimates which are asymptotically exact for the $L^2$ errors of the eigenfunctions. Some numerical results are provided to confirm the theoretical results.

In this work we analyze a postprocessing scheme for improving the guaranteed error bound based on the equilibrated fluxes for the P1 conforming FEM. The improved error bound is shown to be asymptotically exact under suitable conditions on the triangulations and the regularity of the true solution. We also present some numerical results to illustrate the effect of the postprocessing scheme.

The goal of postprocessing in automatic speech recognition is to improve recognition performance by utterance verification at the output of recognition stage. It is focused on the effective rejection of out-of vocabulary words based on the confidence score of hypothesized candidate word. We present two methods for computing confidence scores. Both methods are based on the distance between each observation vector and the representative code vector, which is defined by the most likely code vector at each state. While the first method employs simple time normalization, the second one uses a normalization technique based on the concept of on-line garbage mode[1]. According to the speaker independent isolated words recognition experiment with discrete density HMM, the second method outperforms both the first one and conventional likelihood ratio scoring method[2].

In this paper we discuss some recovery of H(div)-conforming flux approximations from the equilibrated fluxes of Ainsworth and Oden for quadratic finite element methods of second-order elliptic problems. Combined with the hypercircle method of Prager and Synge, these flux approximations lead to a posteriori error estimators which provide guaranteed upper bounds on the numerical error. Furthermore, we prove some superconvergence results for the flux approximations and asymptotic exactness for the error estimator under proper conditions on the triangulation and the exact solution. The results extend those of the previous paper for linear finite element methods.

Human faces have been broadly studied in digital image and video processing fields. An appearance-based method, the adaptive boosting learning algorithm using integral image representations has been successfully employed for face detection, taking advantage of the feature extraction's low computational complexity. In this paper, we propose a face-detection postprocessing method that equalizes instantaneous facial regions in an efficient hardware architecture for use in real-time multimedia applications. The proposed system requires low hardware resources and exhibits robust performance in terms of the movements, zooming, and classification of faces. A series of experimental results obtained using video sequences collected under dynamic conditions are discussed.

Background and Purpose Focal cortical dysplasia (FCD) is one of the most common causes of drug-resistant epilepsy, and necessitates a multimodal evaluation to ensure optimal surgical treatment. This study aimed to determine the supportive value of the morphometric analysis program (MAP) in detecting FCD using data from a single institution in Korea. Methods To develop a standard reference for the MAP, normal-looking MRIs by two scanners that are frequently used in this center were chosen. Patients with drug-resistant epilepsy and FCD after surgery were candidates for the analysis. The three-dimensional T1-weighted MRI scans of the patients were analyzed as test cases using the MAP. Results The MRI scans of 87 patients were included in the analysis. The radiologist detected abnormal findings correlated with FCD (RAD positive [RAD(+)]) in 34 cases (39.1%), while the MAP could detect FCD in 25.3% of cases. A combination of the MAP (MAP[+] cases) with interpretations by the radiologist increased the detection to 42.5% (37 cases). The lesion detection rate was not different according to the type of reference scanners except in one case. MAP(+)/RAD(-) presented in three cases, all of which had FCD type IIa. The detection rate was slightly higher using the same kind of scanner as a reference, but not significantly (35.0% vs. 22.4% p=0.26). Conclusions The results of postprocessing in the MAP for detecting FCD did not depend on the type of reference scanner, and the MAP was the strongest in detecting FCD IIa. We suggested that the MAP could be widely utilized without developing institutional standards and could become an effective tool for detecting FCD lesions.

본 논문에서는 비음수 행렬 분해 기반의 능동 소나 잔향 제거 기법의 성능을 개선하는 연구를 진행하였다. 비음수 행렬 분해 기법은 하나의 비음수 행렬을 두 비음수 행렬의 곱으로 분해하는 기법으로, 잔향 제거 기법에서는 이를 활용하여 입력 신호의 크기 스펙트로그램을 두 행렬 성분, 즉 표적 반향 성분과 잔향 성분으로 분리한다. 기존의 기법에서는 이와 같이 분리된 표적 반향 성분만을 이용하여 원하는 신호를 획득하는 반면, 본 연구에서는 표적 반향 성분과 잔향 성분을 모두 활용하여 잔향 제거 성능을 더욱 향상시키는 후처리 기법에 대해 연구하였으며, 이를 달성하기 위하여 잔향 성분을 평활화하고 위너 마스크를 구성하여 잔향을 제거하는 기법을 제안한다. 본 연구에서는 제안하는 기법의 효과를 검증하기 위하여 시뮬레이션을 진행하였으며, 기존 기법 대비 최대 4 dB의 신호대잡음비 향상을 얻을 수 있었다.

In this paper, we present a study on improving the performance of active sonar reverberation suppression based on Non-negative Matrix Factorization (NMF). NMF is a technique that decomposes a non-negative matrix into the product of two non-negative matrices. In reverberation suppression, this method decomposes the magnitude spectrogram of the input signal into two components: the target echo and the reverberation matrices. While conventional approaches utilize only the separated target echo matrix to recover the desired signal, this study proposes a post-processing method that utilizes both the target echo and reverberation components to enhance reverberation suppression performance. Specifically, the proposed method smooths the reverberation component and constructs a Wiener mask to suppress reverberation. The effectiveness of the proposed method was verified through simulations, demonstrating an improvement in the signal-to-noise ratio of up to 4 dB compared to conventional techniques.

본 논문에서는 기존의 인공신경망을 이용한 자동작곡에서 음악적으로 부족한 부분을 개선하기 위해 조성을 후처리하는 방법과 멜로디에 반복성을 주는 방법 그리고 다수의 곡을 학습하기 위한 평균 신경망 방법을 제안한다. 인공신경망을 이용하여 작곡된 곡의 멜로디는 인공신경망에 학습된 곡의 멜로디에 따라서 출력되는 것으로 음악적으로 특정한 조성에 맞는 곡이 출력되지 않으며 또한 반복적인 멜로디 구성이 나오기 어렵다. 본 논문에서는 이를 해결하기 위하여 인공신경망이 출력한 멜로디를 음악이론에 따라서 특정한 조성으로 후처리하는 방법과 마디구분을 반복적으로 구성하여 멜로디 진행에 반복을 주는 방법을 제안한다. 또한 기존 연구에서 사용한 다수의 곡을 학습하는 방법은 여러 가지 단점이 있었다. 이를 해결하기 위하여 다수의 곡을 학습하는 방법으로 각 곡을 학습한 인공신경망의 가중치를 평균하여 만든 평균 인공신경망을 사용하는 것을 제안한다. 제안한 방법을 적용하여 작곡한 결과 제안한 방법이 기존의 문제점을 해결하는 것을 확인할 수 있었다.

This paper introduces a postprocessing method, an iteration method for melody, and an average neural network method for learning a large number of songs in order to improve musically insufficient parts in automatic composition using existing artificial neural network. The melody of songs composed by artificial neural networks is produced according to the melodies of trained songs, so it can not be a specific tonality and it is difficult to have a repetitive composition. In order to solve these problems, we propose a postprocessing method that converts the melody composed by artificial neural networks into a melody having a specific tonality according to music theory and an iteration method for melody by iteratively composing measure divisions of artificial neural networks. In addition, the existing training method of many songs has some disadvantages. To solve this problem, we adopt an average neural network that is made by averaging the weights of artificial neural networks trained each song. From some experiments, it was confirmed that the proposed method solves the existing problems.

The postprocessing of fingerprint image are widely used to eliminate the false minutiae that caused by skeletonization. This paper presents a new postprocessing algorithm of the skeletonized fingerprint image using SOFM. The proposed postprocessing method showed the good performance for eliminating the spurious minutiae.

In this paper, we proposed a postprocessing algorithm for quantization effects reduction in block coded images using the block classification and adaptive filtering. The proposed method consists of classification, adaptive inter-block filtering, and intra-block filtering. First, each block is classified into one of seven classes based on the characteristics of 8${\times}$8 DCT coefficients. Then each block boundary is filtered by adaptive inter-block filters according to the block classification. Finally for blocks which are classified into edge block, intra-block filtering is peformed. Experimental results show that the proposed method gives better results than the conventional methods from both a subjective and an objective viewpoint.

연속 음성 인식 결과를 자연어 처리 기술과 접목시키기 위해 처리해야 할 두가지 문제점이 있다. 첫째는 말하는 단위와 문서의 띄어쓰기 단위가 일치하지 않는다는 것이고, 둘째는 발음시 형태소 내부 및 형태소 간에 음운 변동 현상이 생긴다는 것이다. 본 논문에서는 이 두가지 문제를 어절생성기와 음절복원기로 해결하고, 생성된 결과들을 형태소 분석하여 실패한 결과들은 교정기를 통해 교정하는 연속 음성 인식 후처리 시스템을 구현하였다. 제안한 시스템의 실험은 두 종류의 음성 말뭉치 ？, 교과서 음성 말뭉치와 사설 음성 말뭉치를 대항으로 수행하였다. 각 말뭉치에 대한 성공률은 각각 93.72%, 92.26% 였고, 이 실험으로 제안한 시스템은 음성 말뭉치의 종류에 민감하지 않는 안정된 시스템임을 알 수 있었다.

There are two problems which will be processed to graft a continuous speech recognition results into natural language processing technique. First, the speaking's unit isn't consistent with text's spacing unit. Second, when it is to be pronounced the phonological alternation phenomena occur inside morphemes or among morphemes. In this paper, we implement the postprocessing system of a continuous speech recognition that above all, solve two problems using the eo-jeol generator and syllable recoveror and morphologically analyze the generated results and then correct the failed results through the corrector. Our system experiments with two kinds of speech corpus, i.e., a primary school text book and editorial corpus. The successful percentage of the former is 93.72%, that of the latter is 92.26%. As results of experiment, we verified that our system is stable regardless the sorts of corpus.

본 논문에서는 통계적 방법을 이용한 후처리기를 설계하고, 구현하여 평가하였다. 통계적인 방법은 처리 속도보다는 공간 효율을 높임으로써 후처리의 성능을 높일 수 있다는 가정에서, 후처리의 성능을 향상시키기 위해서 다음의 3가지 방법을 제안한다. 첫째, 전이 확률에서 중복 표현되는 정보를 정의하고, 제거할 수 있는 방법을 제안한다. 둘째, 정수인 순위값으로부터 실수인 전이 확률의 근사값을 추정해 냄으로써 공간 효율을 높일 수 있는 방법을 제안한다. 셋째, 위의 두가지 방법을 복합적으로 적용하여 공간 효율을 높은 오류 탐지와 오류 교정 방법을 제안한다.

본 논문에서는 블록 기반 DCT (Discrete Cosine Transform)에 의한 압축 영상의 화질을 개선하기 위한 적응적 후처리 기법을 제안한다. 제안하는 알고리즘은 1차원의 비교적 간단한 디블록킹 필터(deblocking filter)와 2차원의 방향성 기반 화질 개선 필터를 연속적으로 적용하는 형태로 이루어져 있다. 먼저, 영상의 국부 통계적 특성에 따라 적응적인 임계값을 설정하여 1차원 디블록킹 필터를 적응적으로 적용한다. 1차원 필터링을 수행한 후 블록간의 연속성을 향상시켜 인지되는 화질을 향상시키기 위하여 업데이트 과정을 수행하게 된다. 이후, 실제 에지(edge) 및 텍스쳐(texture) 정보를 보존하면서 에지 주위의 링잉 현상(ringing effects) 및 블록의 교차점에 존재하는 불연속 현상을 제거하기 위하여 2차원 방향성 기반 화질 개선 필터를 적용한다. 실험 결과 제안한 알고리즘은 기존의 다양한 화질 개선을 위한 후처리 기법에 비해 높은 PSNR을 나타내면서 우수한 주관적 화질 결과를 보인다.

This paper addresses an adaptive postprocessing method applied in the spatial domain for block-based discrete cosine transform (BDCT) coded images. The proposed algorithm is designed by a serial concatenation of a 1D simple smoothing filter and a 2D directional filter. The 1D smoothing filter is applied according to the block type, which is determined by an adaptive threshold. It depends on local statistical properties, and updates block types appropriately by a simple rule, which affects the performance of deblocking processes. In addition, the 2D directional filter is introduced to suppress the ringing effects at the sharp edges and the block discontinuities while preserving true edges and textural information. Comprehensive experiments indicate that the proposed algorithm outperforms many deblocking methods in the literature, in terms of PSNR and subjective visual quality evaluated by GBIM.

병렬 말뭉치 내에서 서로 대응되는 단어를 찾아내는 단어 정렬 작업은 기계 번역에서 가장 기본적으로 수행되는 작업이고 다양한 분야에서 유용하게 사용된다. 본 논문에서는 영한 단어 정렬에서 기존의 통계 기반 정렬 모델의 문제점을 파악하고 이를 해결하기 위해 영한의 품사간 정렬 경향을 단어 정렬에 반영하는 방법을 제안한다. 실험을 통해서 기존 통계 기반 영한 단어 정렬 결과와 비교하여 제안된 방법이 정확률, 재현율, F-measure 측면에서 모두 향상시키는 것을 보였다.