• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.42 no.5 s.305
• /
• pp.123-136
• /
• 2005

We propose a recognition system for superimposed patterns based on selective attention model and SVM which produces better performance than artificial neural network. The proposed selective attention model includes attention layer prior to SVM which affects SVM's input parameters. It also behaves as selective filter. The philosophy behind selective attention model is to find the stopping criteria to stop training and also defines the confidence measure of the selective attention's outcome. Support vector represents the other surrounding sample vectors. The support vector closest to the initial input vector in consideration is chosen. Minimal euclidean distance between the modified input vector based on selective attention and the chosen support vector defines the stopping criteria. It is difficult to define the confidence measure of selective attention if we apply common selective attention model, A new way of doffing the confidence measure can be set under the constraint that each modified input pixel does not cross over the boundary of original input pixel, thus the range of applicable information get increased. This method uses the following information; the Euclidean distance between an input pattern and modified pattern, the output of SVM, the support vector output of hidden neuron that is the closest to the initial input pattern. For the recognition experiment, 45 different combinations of USPS digit data are used. Better recognition performance is seen when selective attention is applied along with SVM than SVM only. Also, the proposed selective attention shows better performance than common selective attention.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.3
• /
• pp.234-240
• /
• 2021

In this paper, the mask-based speech enhancement is improved for effective speech recognition in noise environments. In the mask-based speech enhancement, enhanced spectrum is obtained by multiplying the noisy speech spectrum by the mask. The VoiceFilter (VF) model is used as the mask estimation, and the Spectrogram Inpainting (SI) technique is used to remove residual noise of enhanced spectrum. In this paper, we propose a combined loss to further improve speech enhancement. In order to effectively remove the residual noise in the speech, the positive part of the Triplet loss is used with the component loss. For the experiment TIMIT database is re-constructed using NOISEX92 noise and background music samples with various Signal to Noise Ratio (SNR) conditions. Source to Distortion Ratio (SDR), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligibility (STOI) are used as the metrics of performance evaluation. When the VF was trained with the mean squared error and the SI model was trained with the combined loss, SDR, PESQ, and STOI were improved by 0.5, 0.06, and 0.002 respectively compared to the system trained only with the mean squared error.

• KIPS Transactions on Computer and Communication Systems
• /
• v.10 no.7
• /
• pp.191-198
• /
• 2021

Since the size of training dataset become large and the model is getting deeper to achieve high accuracy in deep learning, the deep neural network training requires a lot of computation and it takes too much time with a single node. Therefore, distributed deep learning is proposed to reduce the training time by distributing computation across multiple nodes. In this study, we propose hybrid allreduce strategy that considers the characteristics of each layer and communication and computational overlapping technique for synchronization of distributed deep learning. Since the convolution layer has fewer parameters than the fully-connected layer as well as it is located at the upper, only short overlapping time is allowed. Thus, butterfly allreduce is used to synchronize the convolution layer. On the other hand, fully-connecter layer is synchronized using ring all-reduce. The empirical experiment results on PyTorch with our proposed scheme shows that the proposed method reduced the training time by up to 33% compared to the baseline PyTorch.

• Journal of the Korean Society of Surveying, Geodesy, Photogrammetry and Cartography
• /
• v.37 no.4
• /
• pp.219-231
• /
• 2019

The specifications of signboards are set for each type of signboards, but the shape and size of the signboard actually installed are not uniform. In addition, because the colors of the signboard are not defined, so various colors are applied to the signboard. Methods for recognizing signboards can be thought of as similar methods of recognizing road signs and license plates, but due to the nature of the signboards, there are limitations in that the signboards can not be recognized in a way similar to road signs and license plates. In this study, we proposed a methodology for recognizing plate-type signboards, which are the main targets of illegal and old signboards, and automatically extracting areas of signboards, using the deep learning-based Faster R-CNN algorithm. The process of recognizing flat type signboards through signboard images captured by using smartphone cameras is divided into two sequences. First, the type of signboard was recognized using deep learning to recognize flat type signboards in various types of signboard images, and the result showed an accuracy of about 71%. Next, when the boundary recognition algorithm for the signboards was applied to recognize the boundary area of the flat type signboard, the boundary of flat type signboard was recognized with an accuracy of 85%.

• The Korean Journal of Applied Statistics
• /
• v.35 no.4
• /
• pp.457-468
• /
• 2022

In aged society, it is important to prevent older people from being disability needing long-term care. The purpose of this study is to develop a prediction model to discover high-risk groups who are likely to be beneficiaries of Long-Term Care Insurance. This study is a retrospective study using database of National Health Insurance Service (NHIS) collected in the past of the study subjects. The study subjects are 7,724,101, the population over 65 years of age registered for medical insurance. To develop the prediction model, we used logistic regression, decision tree, random forest, and multi-layer perceptron neural network. Finally, random forest was selected as the prediction model based on the performances of models obtained through internal and external validation. Random forest could predict about 90% of the older people in need of long-term care using DB without any information from the assessment of eligibility for long-term care. The findings might be useful in evidencebased health management for prevention services and can contribute to preemptively discovering those who need preventive services in older people.

• Korean Journal of Remote Sensing
• /
• v.38 no.5_1
• /
• pp.461-469
• /
• 2022

This paper presents a novel training data extraction approach using semi-supervised learning (SSL)-based classification without the analyst intervention for time-series land-cover mapping. The SSL-based approach first performs initial classification using initial training data obtained from past images including land-cover characteristics similar to the image to be classified. Reliable training data from the initial classification result are then extracted from SSL-based iterative classification using classification uncertainty information and class labels of neighboring pixels as constraints. The potential of the SSL-based training data extraction approach was evaluated from a classification experiment using unmanned aerial vehicle images in croplands. The use of new training data automatically extracted by the proposed SSL approach could significantly alleviate the misclassification in the initial classification result. In particular, isolated pixels were substantially reduced by considering spatial contextual information from adjacent pixels. Consequently, the classification accuracy of the proposed approach was similar to that of classification using manually extracted training data. These results indicate that the SSL-based iterative classification presented in this study could be effectively applied to automatically extract reliable training data for time-series land-cover mapping.

• Journal of the Korean Society of Radiology
• /
• v.16 no.1
• /
• pp.1-6
• /
• 2022

Deep learning is a collection of algorithms that enable learning by summarizing the key contents of large amounts of data; it is being developed to diagnose lesions in the medical imaging field. To evaluate the accuracy of the cerebral hemorrhage diagnosis, we used a convolutional neural network (CNN) to derive the diagnostic accuracy of cerebral parenchyma computed tomography (CT) images and the cerebral parenchyma CT images of areas where cerebral hemorrhages are suspected of having occurred. We compared the accuracy of CNN with different numbers of hidden layers and discovered that CNN with more hidden layers resulted in higher accuracy. The analysis results of the derived CT images used in this study to determine the presence of cerebral hemorrhages are expected to be used as foundation data in studies related to the application of artificial intelligence in the medical imaging industry.

• The Journal of the Acoustical Society of Korea
• /
• v.41 no.3
• /
• pp.326-334
• /
• 2022

In this paper, we propose speech enhancement algorithm as a pre-processing for robust speech recognition in noisy environments. Auxiliary-function-based Independent Vector Analysis (AuxIVA) is performed with weighted covariance matrix using time-varying variances with scaling factor from target masks representing time-frequency contributions of target speech. The mask estimates can be obtained using Neural Network (NN) pre-trained for speech extraction or diffuseness using Coherence-to-Diffuse power Ratio (CDR) to find the direct sounds component of a target speech. In addition, outputs for omni-directional noise are closely chained by sharing the time-varying variances similarly to independent subspace analysis or IVA. The speech extraction method based on AuxIVA is also performed in Independent Low-Rank Matrix Analysis (ILRMA) framework by extending the Non-negative Matrix Factorization (NMF) for noise outputs to Non-negative Tensor Factorization (NTF) to maintain the inter-channel dependency in noise output channels. Experimental results on the CHiME-4 datasets demonstrate the effectiveness of the presented algorithms.

• Journal of Korea Society of Industrial Information Systems
• /
• v.27 no.6
• /
• pp.13-23
• /
• 2022

In this paper, we propose a detailed component image classification algorithm by fashion item for unstructured data retrieval in the fashion field. Due to the COVID-19 environment, AI-based online shopping malls are increasing recently. However, there is a limit to accurate unstructured data search with existing keyword search and personalized style recommendations based on user surfing behavior. In this study, pre-processing using Mask R-CNN was conducted using images crawled from online shopping sites and then classified components for each fashion item through CNN. We obtain the accuaracy for collar of the shirt's as 93.28%, the pattern of the shirt as 98.10%, the 3 classese fit of the jeans as 91.73%, And, we further obtained one for the 4 classes fit of jeans as 81.59% and the color of the jeans as 93.91%. At the results for the decorated items, we also obtained the accuract of the washing of the jeans as 91.20% and the demage of jeans accuaracy as 92.96%.

• Journal of Korea Water Resources Association
• /
• v.54 no.spc1
• /
• pp.1167-1181
• /
• 2021

In relation to the algae bloom, four types of blue-green algae that emit toxic substances are designated and managed as harmful Cyanobacteria, and prediction information using a physical model is being also published. However, as algae are living organisms, it is difficult to predict according to physical dynamics, and not easy to consider the effects of numerous factors such as weather, hydraulic, hydrology, and water quality. Therefore, a lot of researches on algal bloom prediction using machine learning have been recently conducted. In this study, the characteristic importance of water quality factors affecting the occurrence of Cyanobacteria harmful algal blooms (CyanoHABs) were analyzed using the random forest (RF) model for Bohyeonsan Dam and Yeongcheon Dam located in Yeongcheon-si, Gyeongsangbuk-do and also predicted the occurrence of harmful blue-green algae using the machine learning and deep learning models and evaluated their accuracy. The water temperature and total nitrogen (T-N) were found to be high in common, and the occurrence prediction of CyanoHABs using artificial neural network (ANN) also predicted the actual values closely, confirming that it can be used for the reservoirs that require the prediction of harmful cyanobacteria for algal management in the future.