• Proceedings of the Korean Information Science Society Conference
• /
• 1999.10a
• /
• pp.12-14
• /
• 1999

오늘날 멀티미디어 및 인터넷 서비스가 눈에 띄게 증가하면서 다양한 응용분야에서의 동영상 데이터 활용을 급증하였고 이에 사용자가 원하는 동영상 데이터를 빠르고 정확하게 검색하기 위한 내용기반 검색기법이 필수적이다. 본 논문은 high-level features와 더불어 동영상의 고유 내용 속성에 속하는 low-level features를 자동 일반화(generalization)하여 다단계 관리하고 features에 대한 가중치 적용질의를 제공함으로써 기존 내용기반 검색 연구와는 뚜렷한 차별성을 갖는다. 또한 low-level features와 high-level features간의 자동변환(translation)을 가능하게 함으로써 동영상 데이터베이스의 사용자 접근 효율을 한단계 높이고 보다 의미구조화된 동영상 관리 및 내용기반 검색을 지원한다.

• Journal of Electrical Engineering and Technology
• /
• v.13 no.1
• /
• pp.38-50
• /
• 2018

Spatio-temporal load forecasting (STLF) is a foundation for building the prediction-based power map, which could be a useful tool for the visualization and tendency assessment of urban energy application. Constructing one point-forecasting model for each electricity cell in the geographic space is possible; however, it is unadvisable and insufficient, considering the aggregation features of electricity cells and uncertainties in input variables. This paper presents a new STLF method, with a data-driven framework consisting of 3 subroutines: multi-level clustering of cells considering their aggregation features, load regression for each category of cells based on SLS-SVRNs (sparse least squares support vector regression networks), and interval forecasting of spatio-temporal load with sampled blind number. Take some area in Pudong, Shanghai as the region of study. Results of multi-level clustering show that electricity cells in the same category are clustered in geographic space to some extent, which reveals the spatial aggregation feature of cells. For cellular load regression, a comparison has been made with 3 other forecasting methods, indicating the higher accuracy of the proposed method in point-forecasting of spatio-temporal load. Furthermore, results of interval load forecasting demonstrate that the proposed prediction-interval construction method can effectively convey the uncertainties in input variables.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.31B no.6
• /
• pp.104-110
• /
• 1994

We proposed a high speed feature extraction algorithm and developed a feature vector extraction chip for high speed character recognition. It is hard to implement a high speed OCR by software alone with statistical method . Thus, the whole recognition process is divided into functional steps, then pipeline processed so that high speed processing is possible with temporal parallelism of the steps. In this paper we discuss the feature extraction step of the functional steps. To extract feature vector, a character image is normalized to 40$\times$40 pixels. Then, it is divided into 5$\times$5 subregions and 4x4 subregions to construct 41 overlapped subregions(10x10 pixels). It requires to execute more than 500 commands to extract a feature vector of a subregion by software. The proposed algorithm, however, requires only 10 cycles since it can extract a feature vector of a columm of subregion in one cycle with array structure. Thus, it is possible to process 12.000 characters per second with the proposed algorithm. The chip is implemented using EPLD and the effectiveness is proved by developing an OCR using it.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.22 no.6
• /
• pp.793-798
• /
• 2012

In this paper, we explored the new method for extracting feature from the electroencephalography (EEG) signal based on linear regression technique with the orthonormal polynomial bases. At first, EEG signals from electrodes around motor cortex were selected and were filtered in both spatial and temporal filter using band pass filter for alpha and beta rhymic band which considered related to the synchronization and desynchonization of firing neurons population during motor imagery task. Signal from epoch length 1s were fitted into linear regression with Legendre polynomials bases and extract the linear regression weight as final features. We compared our feature to the state of art feature, power band feature in binary classification using support vector machine (SVM) with 5-fold cross validations for comparing the classification accuracy. The result showed that our proposed method improved the classification accuracy 5.44% in average of all subject over power band features in individual subject study and 84.5% of classification accuracy with forward feature selection improvement.

• ETRI Journal
• /
• v.41 no.2
• /
• pp.235-241
• /
• 2019

In this paper, we propose a rank-weighted reconstruction feature to improve the robustness of a feed-forward deep neural network (FFDNN)-based acoustic model. In the FFDNN-based acoustic model, an input feature is constructed by vectorizing a submatrix that is created by slicing the feature vectors of frames within a context window. In this type of feature construction, the appropriate context window size is important because it determines the amount of trivial or discriminative information, such as redundancy, or temporal context of the input features. However, we ascertained whether a single parameter is sufficiently able to control the quantity of information. Therefore, we investigated the input feature construction from the perspectives of rank and nullity, and proposed a rank-weighted reconstruction feature herein, that allows for the retention of speech information components and the reduction in trivial components. The proposed method was evaluated in the TIMIT phone recognition and Wall Street Journal (WSJ) domains. The proposed method reduced the phone error rate of the TIMIT domain from 18.4% to 18.0%, and the word error rate of the WSJ domain from 4.70% to 4.43%.

• Proceedings of the KSRS Conference
• /
• 2003.11a
• /
• pp.1346-1348
• /
• 2003

In this paper, the interferometric coherence was estimated and the amplitude intensity was extracted using the repeat-pass interferometric data, acquired by European Remote Sensing Satellite 1 and 2. Then discrimination and classification of surface land types in Zhangjiakou test site, Hebei Province were carried out based on the coherence estimation and the intensity extraction. Seven types of land were discriminated and classified, including in two different types of meadows, woodland, dry land, grassland, steppe and water body. The backscatter and coherence characteristics of these land types on the multi-temporal images were analyzed, and the change of surface features with time series was also discussed.

• Proceedings of the Korean Information Science Society Conference
• /
• 2012.06c
• /
• pp.406-407
• /
• 2012

Texture transfer is a NPR technique for expressing various styles according to source (reference) image. By late 2000s, there are many texture transfer researches. But video base researchers are not active. Moreover, they didn't use important feature like directional information which need to express detail characteristics of target. So, we propose a new method to generate texture transfer animation (using video) with directional effect for maintaining temporal coherence and controlling coherence direction of texture. For maintaining temporal coherence, we use optical flow and confidence map to adapt for occlusion/disocclusion boundaries. And we control direction of texture for taking structure of input. For expressing various texture effects according to different regions, we calculate gradient based on directional weight. With these techniques, our algorithm can make animation result that maintain temporal coherence and express directional texture effect. It is reflect the characteristics of source and target image well. And our result can express various texture directions automatically.

• Journal of Information Processing Systems
• /
• v.17 no.2
• /
• pp.242-252
• /
• 2021

Following the intuition that the local information in time instances is hardly incorporated into the posterior sequence in long short-term memory (LSTM), this paper proposes an attention augmented mechanism for fault diagnosis of the complex chemical process data. Unlike conventional fault diagnosis and classification methods, an attention mechanism layer architecture is introduced to detect and focus on local temporal information. The augmented deep network results preserve each local instance's importance and contribution and allow the interpretable feature representation and classification simultaneously. The comprehensive comparative analyses demonstrate that the developed model has a high-quality fault classification rate of 95.49%, on average. The results are comparable to those obtained using various other techniques for the Tennessee Eastman benchmark process.

• Journal of Broadcast Engineering
• /
• v.24 no.5
• /
• pp.735-742
• /
• 2019

In this paper, we propose a method of extracting speech features for phoneme recognition based on spikegram. The Fourier-transform-based features are widely used in phoneme recognition, but they are not extracted in a biologically plausible way and cannot have high temporal resolution due to the frame-based operation. For better phoneme recognition, therefore, it is desirable to have a new method of extracting speech features, which analyzes speech signal in high temporal resolution following the model of human auditory system. In this paper, we analyze speech signal based on a spikegram that models feature extraction and transmission in auditory system, and then propose a method of feature extraction from the spikegram for phoneme recognition. We evaluate the performance of proposed features by using a DNN-based phoneme recognizer and confirm that the proposed features provide better performance than the Fourier-transform-based features for short-length phonemes. From this result, we can verify the feasibility of new speech features extracted based on auditory model for phoneme recognition.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.11 no.3
• /
• pp.1670-1683
• /
• 2017

Mitotic event recognition is a crucial and challenging task in biomedical applications. In this paper, we introduce the slow feature analysis and propose a fully-automated mitotic event recognition method for cell populations imaged with time-lapse phase contrast microscopy. The method includes three steps. First, a candidate sequence extraction method is utilized to exclude most of the sequences not containing mitosis. Next, slow feature is learned from the candidate sequences using slow feature analysis. Finally, a hidden conditional random field (HCRF) model is applied for the classification of the sequences. We use a supervised SFA learning strategy to learn the slow feature function because the strategy brings image content and discriminative information together to get a better encoding. Besides, the HCRF model is more suitable to describe the temporal structure of image sequences than nonsequential SVM approaches. In our experiment, the proposed recognition method achieved 0.93 area under curve (AUC) and 91% accuracy on a very challenging phase contrast microscopy dataset named C2C12.