• Proceedings of the Korean Information Science Society Conference
• /
• 2011.06c
• /
• pp.373-376
• /
• 2011

얼굴 영상에서 구성요소(눈썹, 눈, 코, 입 등)의 존재에 따라 보는 사람의 얼굴 인식 정확도는 큰 영향을 받는다. 이는 인간의 뇌에서 얼굴 정보를 처리하는 과정은 얼굴 전체 영역 뿐만 아니라, 부분적인 얼굴 구성요소의 특징들도 고려함을 말한다. 비음수 행렬 분해(NMF: Non-negative Matrix Factorization)는 이러한 얼굴 영역에서 부분적인 특징들을 잘 표현하는 기저영상들을 찾아내는데 효과적임을 보여주었으나, 각 기저영상들의 중요도는 알 수 없었다. 본 논문에서는 NMF로 찾아진 기저영상들에 대응되는 인코딩 정보를 SLR(Sparse Logistic Regression)을 이용하여 성별 인식에 중요한 부분 영역들을 찾고자 한다. 실험에서는 주성분분석(PCA)과 비교를 통해 NMF를 이용한 기저영상 및 특징 벡터 추출이 좋은 성능을 보여주고, 대표적 이진 분류 알고리즘인 SVM(Support Vector Machine)과 비교를 통해 SLR을 이용한 특징 벡터 선택이 나은 성능을 보여줌을 확인하였다. 또한 SLR로 확인된 각 기저영상에 대한 가중치를 통하여 인식 과정에서 중요한 얼굴 영역들을 확인할 수 있다.

• Proceedings of the Korea Contents Association Conference
• /
• 2018.05a
• /
• pp.485-486
• /
• 2018

In this paper, we consider a binary recovery framework of the Compressed Sensing (CS) problem. We derive an upper bound for $L_0$ recovery performance of a binary sparse signal in terms of the dimension N and sparsity K of signals, the number of measurements M. We show that the upper bound obtained from this work goes to the limit bound when the sensing matrix sufficiently become dense. In addition, for perfect recovery performance, if the signals are very sparse, the sensing matrices required for $L_0$ recovery are little more dense.

• The Transactions of the Korean Institute of Electrical Engineers C
• /
• v.49 no.1
• /
• pp.52-58
• /
• 2000

Signal propagation in nonlinear optical fiber is analyzed numerically by using SS-FEM (Split-Step Finite Element Method). By adopting cubic element function in FEM, soliton equation of which exact solution was well known, has been solved. Also, accuracy of numerical results and computing times are compared with those of Fourier method, and we have found that solution obtained from using FEM was very relatively accurate. Especially, to reduce CPU time in matrix computation in each step, the matrix imposed by the boundary condition is approximated as a sparse matrix. As a result, computation time was shortened even with the same or better accuracy when compared to those of the conventional FEM and Fourier method.

• Communications of the Korean Mathematical Society
• /
• v.11 no.4
• /
• pp.1175-1185
• /
• 1996

We study the computation of sparse null bases of equilibrium matrices in the context of structural optimization and incompressible fluid flow. In our approach we emphasize the parallel computatin and examine the applications. New block decomposition and node ordering schemes are suggested, and numerical examples are considered.

• KIISE Transactions on Computing Practices
• /
• v.20 no.11
• /
• pp.610-614
• /
• 2014

Like many types of scientific data, results from simulations of volcanic ash diffusion are of a clustered sparse matrix in the netCDF format. Since these data sets are large in size, they generate high storage and transmission costs. In this paper, we suggest a new method that reduces the size of the data of volcanic ash diffusion simulations by converting the multi-dimensional index to a single dimension and keeping only the starting point and length of the consecutive zeros. This method presents performance that is almost as good as that of ZIP format compression, but does not destroy the netCDF structure. The suggested method is expected to allow for storage space to be efficiently used by reducing both the data size and the network transmission time.

• Journal of Korea Society of Digital Industry and Information Management
• /
• v.17 no.4
• /
• pp.53-62
• /
• 2021

Deep learning models such as convolutional neural networks and recurrent neual networks process a huge amounts of data, so they require a lot of storage and consume a lot of time and power due to memory access. Recently, research is being conducted to reduce memory usage and access by compressing data using the feature that many of deep learning data are highly sparse and localized. In this paper, we propose a compression-decompression method of storing only the non-zero data and the location information of the non-zero data excluding zero data. In order to make the location information of non-zero data, the matrix data is divided into sections uniformly. And whether there is non-zero data in the corresponding section is indicated. In this case, section division is not executed only once, but repeatedly executed, and location information is stored in each step. Therefore, it can be properly compressed according to the ratio and distribution of zero data. In addition, we propose a hardware structure that enables compression and decompression without complex operations. It was designed and verified with Verilog, and it was confirmed that it can be used in hardware deep learning accelerators.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.39A no.12
• /
• pp.739-745
• /
• 2014

It has been shown in compressive sensing that every s-sparse $x{\in}R^N$ can be recovered from the measurement vector y=Ax or the noisy vector y=Ax+e via ${\ell}_1$-minimization as soon as the 3s-restricted isometry constant of the sensing matrix A is smaller than 1/2 or smaller than $1/\sqrt{3}$ by applying the Iterative Hard Thresholding (IHT) algorithm. However, recovery can be guaranteed by practical algorithms for some certain assumptions of acquisition schemes. One of the key assumption is that the sensing matrix must satisfy the Restricted Isometry Property (RIP), which is often violated in the setting of many practical applications. In this paper, we studied a generalization of RIP, called Restricted Biorthogonality Property (RBOP) for anisotropic cases, and the new recovery algorithms called oblique pursuits. Then, we provide an analysis on the success of sparse recovery in terms of restricted biorthogonality constant for the IHT algorithms.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.7
• /
• pp.3594-3607
• /
• 2017

The traditional feature extraction methods such as principal component analysis (PCA) cannot obtain the local structure of the samples, and locally linear embedding (LLE) cannot obtain the global structure of the samples. However, a common drawback of existing PCA and LLE algorithm is that they cannot deal well with the sparse problem of the samples. Therefore, by integrating the globality of PCA and the locality of LLE with a sparse constraint, we developed an improved and unsupervised difference algorithm called Sparse Difference Embedding (SDE), for dimensionality reduction of high-dimensional data in small sample size problems. Significantly differing from the existing PCA and LLE algorithms, SDE seeks to find a set of perfect projections that can not only impact the locality of intraclass and maximize the globality of interclass, but can also simultaneously use the Lasso regression to obtain a sparse transformation matrix. This characteristic makes SDE more intuitive and more powerful than PCA and LLE. At last, the proposed algorithm was estimated through experiments using the Yale and AR face image databases and the USPS handwriting digital databases. The experimental results show that SDE outperforms PCA LLE and UDP attributed to its sparse discriminating characteristics, which also indicates that the SDE is an effective method for face recognition.

• Journal of Mechanical Science and Technology
• /
• v.15 no.8
• /
• pp.1090-1096
• /
• 2001

In this paper, new methods for efficiently solving linear acceleration equations of multibody dynamic simulation exploiting sparsity for real-time simulation are presented. The coefficient matrix of the equations tends to have a large number of zero entries according to the relative joint coordinate numbering. By adequate joint coordinate numbering, the matrix has minimum off-diagonal terms and a block pattern of non-zero entries and can be solved efficiently. The proposed methods, using sparse Cholesky method and recursive block mass matrix method, take advantages of both the special structure and the sparsity of the coefficient matrix to reduce computation time. The first method solves the η$\times$η sparse coefficient matrix for the accelerations, where η denotes the number of relative coordinates. In the second method, for vehicle dynamic simulation, simple manipulations bring the original problem of dimension η$\times$η to an equivalent problem of dimension 6$\times$6 to be solved for the accelerations of a vehicle chassis. For vehicle dynamic simulation, the proposed solution methods are proved to be more efficient than the classical approaches using reduced Lagrangian multiplier method. With the methods computation time for real-time vehicle dynamic simulation can be reduced up to 14 per cent compared to the classical approach.

• Journal of KIISE:Software and Applications
• /
• v.35 no.4
• /
• pp.255-264
• /
• 2008

This paper proposes a novel method using K-means and Non-negative matrix factorization (NMF) for topic -based multi-document summarization. NMF decomposes weighted term by sentence matrix into two sparse non-negative matrices: semantic feature matrix and semantic variable matrix. Obtained semantic features are comprehensible intuitively. Weighted similarity between topic and semantic features can prevent meaningless sentences that are similar to a topic from being selected. K-means clustering removes noises from sentences so that biased semantics of documents are not reflected to summaries. Besides, coherence of document summaries can be enhanced by arranging selected sentences in the order of their ranks. The experimental results show that the proposed method achieves better performance than other methods.