• Journal of Digital Convergence
• /
• v.19 no.10
• /
• pp.253-263
• /
• 2021

Machine learning is the process of constructing a cost function using learning data used for learning and an artificial neural network to predict the data, and finding parameters that minimize the cost function. Parameters are changed by using the gradient-based method of the cost function. The more complex the digital signal and the more complex the problem to be learned, the more complex and deeper the structure of the artificial neural network. Such a complex and deep neural network structure can cause over-fitting problems. In order to avoid over-fitting, a weight decay regularization method of parameters is used. We additionally use the value of the cost function in this method. In this way, the accuracy of machine learning is improved, and the superiority is confirmed through numerical experiments. These results derive accurate values for a wide range of artificial intelligence data through machine learning.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.10a
• /
• pp.391-393
• /
• 2021

In this paper, we introduce a long short-term memory (LSTM)-based fall detection system using TensorFlow that can detect falls occurring in the elderly in daily living. 3-axis accelerometer data are aggregated for fall detection, and then three types of parameter are calculated. 4 types of activity of daily living (ADL) and 3 types of fall situation patterns are classified. The parameterized data applied to LSTM. Learning proceeds until the Loss value becomes 0.5 or less. The results are calculated for each parameter θ, SVM, and GSVM. The best result was GSVM, which showed Sensitivity 98.75%, Specificity 99.68%, and Accuracy 99.28%.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.43 no.2 s.308
• /
• pp.1-11
• /
• 2006

In this paper, we propose the novel video retrieval algorithm using non-parametric based motion classification in the shot-based video indexing structure. The proposed system firstly gets the key frame and motion information from each shot segmented by scene change detection method, and then extracts visual features and non-parametric based motion information from them. Finally, we construct real-time retrieval system supporting similarity comparison of these spatio-temporal features. After the normalized motion vector fields is created from MPEG compressed stream, the extraction of non-parametric based motion feature is effectively achieved by discretizing each normalized motion vectors into various angle bins, and considering a mean, a variance, and a direction of these bins. We use the edge-based spatial descriptor to extract the visual feature in key frames. Experimental evidence shows that our algorithm outperforms other video retrieval methods for image indexing and retrieval. To index the feature vectors, we use R*-tree structures.

• Journal of Broadcast Engineering
• /
• v.8 no.1
• /
• pp.19-29
• /
• 2003

The demand for high-resolution images is gradually increasing, whereas many imaging systems yield aliased and undersampled images during image acquisition. In this paper, we propose a high-resolution image reconstruction algorithm considering inaccurate subpixel registration. A regularized Iterative reconstruction algorithm is adopted to overcome the ill-posedness problem resulting from inaccurate subpixel registration. In particular, we use multichannel image reconstruction algorithms suitable for application with multiframe environments. Since the registration error in each low-resolution has a different pattern, the regularization parameters are determined adaptively for each channel. We propose a methods for estimating the regularization parameter automatically. The preposed algorithm are robust against the registration error noise. and they do not require any prior information about the original image or the registration error process. Experimental results indicate that the proposed algorithms outperform conventional approaches in terms of both objective measurements and visual evaluation.

• Annual Conference on Human and Language Technology
• /
• 1999.10e
• /
• pp.184-190
• /
• 1999

본 논문에서는 Text-to-Speech 시스템에서 사용할 억양 모델을 위해 음성 DB에서 모델 파라미터와 피치 검색테이블(lookup-table)을 추출하여 미리 구성하고, 합성시에는 이를 추정하여 최종 F0 값을 생성하는 자료기반 접근방식(data-driven approach)을 사용한다. 어절 경계강도(break-index)는 경계강도의 특성에 따라 고정적 경계강도와 가변적 경계강도로 세분화하여 사용하였고, 예측된 경계강도를 기준으로 억양구(Intonation Phrase)와 액센트구(Accentual Phrase)를 설정하였다. 특히, 액센트구 모델은 인지적, 음향적으로 중요한 정점(peak)을 정확하게 모델링하는 것에 주안점을 두어 정점(peak)의 시간축, 주파수축 값과 이를 기준으로 한 앞뒤 기울기를 추정하여 4개의 파라미터로 설정하였고, 이 파라미터들은 CART(Classification and Regression Tree)를 이용하여 예측규칙을 만들었다. 경계음조가 나타나는 조사, 어미는 정규화된(normalized) 피치값과 key-index로 구성되는 검색테이블을 만들어 보다 정교하게 피치값을 예측하였다. 본 논문에서 제안한 억양 모델을 본 연구실에서 제작한 음성합성기를 통해 합성하여 청취실험을 거친 결과, 기존의 상용 Text-to-Speech 시스템에 비해 자연스러운 합성음을 얻을 수 있었다.

• Journal of the Korea Society of Computer and Information
• /
• v.15 no.12
• /
• pp.125-132
• /
• 2010

In vocabulary recognition has unseen tri-phone appeared when recognition training. This system has not been created beginning estimation figure of model parameter. It's bad points could not be created that model for phoneme data. Therefore it's could not be secured accuracy of Gaussian model. To improve suggested Gaussian model to optimized method of model parameter using probability distribution. To improved of confidence that Gaussian model to optimized of probability distribution to offer by accuracy and to support searching of phoneme data. This paper suggested system performance comparison as a result of recognition improve represent 1.7% by out-of vocabulary rejection algorithm using normalization confidence.

• Journal of the Korea Computer Industry Society
• /
• v.9 no.1
• /
• pp.21-28
• /
• 2008

In this paper, the edge parameters were drawn from speaking image around lip and effective automatic lip reading system to recognize the Korean 'a/e/i/o/u' 5 owels were constructed using the parameter. Speaking images around lip were divided into $5{\times}5$ pane. In each pane the number of digital edge element using Sobel operator were evaluated. The observational error between samples was corrected by using normalization method and the normalized value is used for parameter In the experiment to convince the strength of parameter, 50 normal persons were sampled. The images of 10 persons were analyzed and the images of another 40 persons were experimented for recognition. 500 data are gathered and analyzed. Based on this analysis, the neural net system is constructed and the recognition experiments are performed for 400 data. The neural net system gave the best recognition result of 91.1%.

• Proceedings of the Korea Multimedia Society Conference
• /
• 2000.11a
• /
• pp.271-276
• /
• 2000

본 논문은 칼라 성분들간의 차분 영상과 휘도 영상을 이용하여 산출한 색차 휘도합 영상을 대상으로 블록에 기반한 영상 분할을 수행하여 객체의 형상 정보를 추출함으로써 분할 특성을 개선한 블록 기반 칼라 영상 분할 기법에 관한 것이다. 우선, R, G, B 영상들 간의 차분 성분들을 구하여 합산한 후, 이를 정규화하여 색차합 영상을 구한다. 다음으로 화소 단위로 휘도 영상의 상위 2비트와 정하화된 색차합 영상의 하위 6비트를 결합하여 색차 휘도합 영상을 얻는다. 이후, 기설정된 크기의 블록으로 분할된 색차 휘도합 영상의 각 블록을 질감 블록과 단순 블록 및 에지 블록으로 분류하고 각 유형의 블록별로 병합한 후, 기설정된 마커 배정 규칙에 따라 선택적으로 마커를 부여한다. 마지막으로, 마커가 부여되지 않은 블록을 대상으로 화소 단위의 워터쉐드 알고리즘을 적용함으로써 자연스러운 형상 정보를 얻을 수 있다. 컴퓨터 시뮬레이션 결과를 통해 고찰할 때, 제안된 방범은 질감 영역에서의 과분할의 문제와 과도한 연산량의 부담을 효과적으로 경감시킬 수 있으나, 더불어, 영상 분할용 파라미터들의 민감도가 낮아 서로 다른 화소 분포 특성온 갖는 영상들에 전역적인 파라미터들사용할 수 있을 뿐만 아니라 특히, 색차 휘도합 영상에 반영된 색차 성분에 힘입어 저대조 경계면에서의 분할 특성을 현저히 개선시킬 수 있는 이점이 있다.

• Journal of KIISE
• /
• v.44 no.7
• /
• pp.668-673
• /
• 2017

Deep neural networks have shown remarkable performance in various fields of pattern recognition such as voice recognition, image recognition and object detection. However, underlying mechanisms of the network have not been fully revealed. In this paper, we focused on empirical analysis of the network parameters. The Faster R-CNN(region-based convolutional neural network) was used as a baseline network of our work and three important parameters were analyzed: the dropout ratio which prevents the overfitting of the neural network, the size of the anchor boxes and the activation function. We also compared the performance of dropout and batch normalization. The network performed favorably when the dropout ratio was 0.3 and the size of the anchor box had not shown notable relation to the performance of the network. The result showed that batch normalization can't entirely substitute the dropout method. The used leaky ReLU(rectified linear unit) with a negative domain slope of 0.02 showed comparably good performance.

• The KIPS Transactions:PartB
• /
• v.14B no.1 s.111
• /
• pp.5-12
• /
• 2007

This paper gives a novel approach that can be represented an image efficiently with its region and shape information together. To do this, we introduced a region-centralized shape descriptor(RCSD) that the size of region only exists at a center point. RCSD consists of circles with three parameters, the distance and the angle between the tenter points, and the diameter, respectively We verified the RCSD parameters to have an information of shape. We can be proved this by reconstructing the shape from the given parameters and evaluated the difference between the its image and the original one. To get this image, we find the estimated points on the contour from the parameters, and connect them by using an interpolation. According to the evaluation, we obtained 88% performance for real images, and showed that it can be used efficiently for representing the binary images. Also we cu make RCSD parameters to be the normalized patterns to have an invariant of its scale or position, and expand them to improve the quality of the performance.