• Journal of Intelligence and Information Systems
• /
• v.27 no.2
• /
• pp.17-32
• /
• 2021

Speaker recognition is generally divided into speaker identification and speaker verification. Speaker recognition plays an important function in the automatic voice system, and the importance of speaker recognition technology is becoming more prominent as the recent development of portable devices, voice technology, and audio content fields continue to expand. Previous speaker recognition studies have been conducted with the goal of automatically determining who the speaker is based on voice files and improving accuracy. Speech is an important sociolinguistic subject, and it contains very useful information that reveals the speaker's attitude, conversation intention, and personality, and this can be an important clue to speaker recognition. The final ending used in the speaker's speech determines the type of sentence or has functions and information such as the speaker's intention, psychological attitude, or relationship to the listener. The use of the terminating ending has various probabilities depending on the characteristics of the speaker, so the type and distribution of the terminating ending of a specific unidentified speaker will be helpful in recognizing the speaker. However, there have been few studies that considered speech in the existing text-based speaker recognition, and if speech information is added to the speech signal-based speaker recognition technique, the accuracy of speaker recognition can be further improved. Hence, the purpose of this paper is to propose a novel method using speech style expressed as a sentence-final ending to improve the accuracy of Korean speaker recognition. To this end, a method called sentence sequencing that generates vector values by using the type and frequency of the sentence-final ending appearing in the utterance of a specific person is proposed. To evaluate the performance of the proposed method, learning and performance evaluation were conducted with a actual drama script. The method proposed in this study can be used as a means to improve the performance of Korean speech recognition service.

• Korean Journal of Agricultural and Forest Meteorology
• /
• v.23 no.4
• /
• pp.251-267
• /
• 2021

Remotely sensed vegetation indices (VIs) are empirically related with gross primary productivity (GPP) in various spatio-temporal scales. The uncertainties in GPP-VI relationship increase with temporal resolution. Uncertainty also exists in the eddy covariance (EC)-based estimation of GPP, arising from the partitioning of the measured net ecosystem CO2 exchange (NEE) into GPP and ecosystem respiration (RE). For two forests and two agricultural sites, we correlated the EC-derived GPP in various time scales with three different near-surface remotely sensed VIs: (1) normalized difference vegetation index (NDVI), (2) enhanced vegetation index (EVI), and (3) near infrared reflectance from vegetation (NIRv) along with NIRvP (i.e., NIRv multiplied by photosynthetically active radiation, PAR). Among the compared VIs, NIRvP showed highest correlation with half-hourly and monthly GPP at all sites. The NIRvP was used to test the reliability of GPP derived by two different NEE partitioning methods: (1) original KoFlux methods (GPP_Ori) and (2) machine-learning based method (GPP_ANN). GPP_ANN showed higher correlation with NIRvP at half-hourly time scale, but there was no difference at daily time scale. The NIRvP-GPP correlation was lower under clear sky conditions due to co-limitation of GPP by other environmental conditions such as air temperature, vapor pressure deficit and soil moisture. However, under cloudy conditions when photosynthesis is mainly limited by radiation, the use of NIRvP was more promising to test the credibility of NEE partitioning methods. Despite the necessity of further analyses, the results suggest that NIRvP can be used as the proxy of GPP at high temporal-scale. However, for the VIs-based GPP estimation with high temporal resolution to be meaningful, complex systems-based analysis methods (related to systems thinking and self-organization that goes beyond the empirical VIs-GPP relationship) should be developed.

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1181-1189
• /
• 2022

Clouds are composed of tiny water droplets, ice crystals, or mixtures suspended in the atmosphere and cover about two-thirds of the Earth's surface. Cloud detection in satellite images is a very difficult task to separate clouds and non-cloud areas because of similar reflectance characteristics to some other ground objects or the ground surface. In contrast to thick clouds, which have distinct characteristics, thin transparent clouds have weak contrast between clouds and background in satellite images and appear mixed with the ground surface. In order to overcome the limitations of transparent clouds in cloud detection, this study conducted cloud detection focusing on transparent clouds using machine learning techniques (Random Forest [RF], Convolutional Neural Networks [CNN]). As reference data, Cloud Mask and Cirrus Mask were used in MOD35 data provided by MOderate Resolution Imaging Spectroradiometer (MODIS), and the pixel ratio of training data was configured to be about 1:1:1 for clouds, transparent clouds, and clear sky for model training considering transparent cloud pixels. As a result of the qualitative comparison of the study, bothRF and CNN successfully detected various types of clouds, including transparent clouds, and in the case of RF+CNN, which mixed the results of the RF model and the CNN model, the cloud detection was well performed, and was confirmed that the limitations of the model were improved. As a quantitative result of the study, the overall accuracy (OA) value of RF was 92%, CNN showed 94.11%, and RF+CNN showed 94.29% accuracy.

• Journal of Korea Water Resources Association
• /
• v.54 no.spc1
• /
• pp.1095-1105
• /
• 2021

The advancement of dam operation is further required due to the upcoming rainy season, typhoons, or torrential rains. Besides, physical models based on specific rules may sometimes have limitations in controlling the release discharge of dam due to inherent uncertainty and complex factors. This study aims to forecast the water level of the nearest station to the dam multi-timestep-ahead and evaluate the availability when it makes a decision for a release discharge of dam based on LSTM (Long Short-Term Memory) of deep learning. The LSTM model was trained and tested on eight data sets with a 1-hour temporal resolution, including primary data used in the dam operation and downstream water level station data about 13 years (2009~2021). The trained model forecasted the water level time series divided by the six lead times: 1, 3, 6, 9, 12, 18-hours, and compared and analyzed with the observed data. As a result, the prediction results of the 1-hour ahead exhibited the best performance for all cases with an average accuracy of MAE of 0.01m, RMSE of 0.015 m, and NSE of 0.99, respectively. In addition, as the lead time increases, the predictive performance of the model tends to decrease slightly. The model may similarly estimate and reliably predicts the temporal pattern of the observed water level. Thus, it is judged that the LSTM model could produce predictive data by extracting the characteristics of complex hydrological non-linear data and can be used to determine the amount of release discharge from the dam when simulating the operation of the dam.

• Journal of Korean Tunnelling and Underground Space Association
• /
• v.24 no.3
• /
• pp.247-262
• /
• 2022

In domestic tunnels, it is mandatory to install CCTVs in tunnels longer than 200 m which are also recommended by installation of a CCTV-based automatic accident detection system. In general, the CCTVs in the tunnel are installed at a low height as well as near by the moving vehicles due to the spatial limitation of tunnel structure, so a severe perspective effect takes place in the distance of installed CCTV and moving vehicles. Because of this effect, conventional CCTV-based accident detection systems in tunnel are known in general to be very hard to achieve the performance in detection of unexpected accidents such as stop or reversely moving vehicles, person on the road and fires, especially far from 100 m. Therefore, in this study, the region of interest is set up and a new concept of inverse perspective transformation technique is introduced. Since moving vehicles in the transformed image is enlarged proportionally to the distance from CCTV, it is possible to achieve consistency in object detection and identification of actual speed of moving vehicles in distance. To show this aspect, two datasets in the same conditions are composed with the original and the transformed images of CCTV in tunnel, respectively. A comparison of variation of appearance speed and size of moving vehicles in distance are made. Then, the performances of the object detection in distance are compared with respect to the both trained deep-learning models. As a result, the model case with the transformed images are able to achieve consistent performance in object and accident detections in distance even by 200 m.

• Korean Journal of Cognitive Science
• /
• v.33 no.2
• /
• pp.109-133
• /
• 2022

Episodic memory consists of a core event and the associated contexts. Although the role of the hippocampus and its neighboring regions in contextual representations during encoding has become increasingly evident, it remains unclear how these regions handle various context-specific information other than spatio-temporal contexts. Using high-resolution functional MRI, we explored the patterns of the medial temporal lobe (MTL) and cortical regions' involvement during the encoding of various types of contextual information (i.e., journalism principle 5W1H): "Who did it?," "Why did it happen?," "What happened?," "When did it happen?," "Where did it happen?," and "How did it happen?" Participants answered six different contextual questions while looking at simple experimental events consisting of two faces with one object on the screen. The MTL was divided to sub-regions by hierarchical clustering from resting-state data. General linear model analyses revealed a stronger activation of MTL sub-regions, the prefrontal lobe (PFC), and the inferior parietal lobule (IPL) during social (Who), behavioral (How), and intentional (Why) contextual processing when compared with spatio-temporal (Where/When) contextual processing. To further investigate the functional networks involved in contextual encoding dissociation, a multivariate pattern analysis was conducted with features selected as the task-based connectivity links between the hippocampal subfields and PFC/IPL. Each social, behavioral, and intentional contextual processing was individually and successfully classified from spatio-temporal contextual processing, respectively. Thus, specific contexts in episodic memory, namely social, behavior, and intention, involve distinct functional connectivity patterns that are distinct from those for spatio-temporal contextual memory.

• Journal of Intelligence and Information Systems
• /
• v.29 no.1
• /
• pp.143-174
• /
• 2023

With the development of deep learning technologies, Artificial Intelligence powered Optical Character Recognition (AI-OCR) has evolved to read multiple languages from various forms of images accurately. For the financial industry, where a large number of diverse documents are processed through manpower, the potential for using AI-OCR is great. In this study, we present a configuration and a design of an AI-OCR modality for use in the financial industry and discuss the platform construction with application cases. Since the use of financial domain data is prohibited under the Personal Information Protection Act, we developed a deep learning-based data generation approach and used it to train the AI-OCR models. The AI-OCR models are trained for image preprocessing, text recognition, and language processing and are configured as a microservice architected platform to process a broad variety of documents. We have demonstrated the AI-OCR platform by applying it to financial domain tasks of document sorting, document verification, and typing assistance The demonstrations confirm the increasing work efficiency and conveniences.

• Korean Journal of Remote Sensing
• /
• v.39 no.2
• /
• pp.207-221
• /
• 2023

This study suggests deep neural network models for estimating air temperature with Level 1B (L1B) datasets of GEO-KOMPSAT-2A (GK-2A). The temperature at 1.5 m above the ground impact not only daily life but also weather warnings such as cold and heat waves. There are many studies to assume the air temperature from the land surface temperature (LST) retrieved from satellites because the air temperature has a strong relationship with the LST. However, an algorithm of the LST, Level 2 output of GK-2A, works only clear sky pixels. To overcome the cloud effects, we apply a deep neural network (DNN) model to assume the air temperature with L1B calibrated for radiometric and geometrics from raw satellite data and compare the model with a linear regression model between LST and air temperature. The root mean square errors (RMSE) of the air temperature for model outputs are used to evaluate the model. The number of 95 in-situ air temperature data was 2,496,634 and the ratio of datasets paired with LST and L1B show 42.1% and 98.4%. The training years are 2020 and 2021 and 2022 is used to validate. The DNN model is designed with an input layer taking 16 channels and four hidden fully connected layers to assume an air temperature. As a result of the model using 16 bands of L1B, the DNN with RMSE 2.22℃ showed great performance than the baseline model with RMSE 3.55℃ on clear sky conditions and the total RMSE including overcast samples was 3.33℃. It is suggested that the DNN is able to overcome cloud effects. However, it showed different characteristics in seasonal and hourly analysis and needed to append solar information as inputs to make a general DNN model because the summer and winter seasons showed a low coefficient of determinations with high standard deviations.

• Korean Journal of Remote Sensing
• /
• v.39 no.6_1
• /
• pp.1283-1297
• /
• 2023

Mid-wave infrared (MWIR) imagery, due to its ability to capture the temperature of land cover and objects, serves as a crucial data source in various fields including environmental monitoring and defense. The KOMPSAT-3A satellite acquires MWIR imagery with high spatial resolution compared to other satellites. However, the limited spatial resolution of MWIR imagery, in comparison to electro-optical (EO) imagery, constrains the optimal utilization of the KOMPSAT-3A data. This study aims to create a highly visible MWIR fusion image by leveraging the edge information from the KOMPSAT-3A panchromatic (PAN) image. Preprocessing is implemented to mitigate the relative geometric errors between the PAN and MWIR images. Subsequently, we employ a pre-trained pixel difference network (PiDiNet), a deep learning-based edge information extraction technique, to extract the boundaries of objects from the preprocessed PAN images. The MWIR fusion imagery is then generated by emphasizing the brightness value corresponding to the edge information of the PAN image. To evaluate the proposed method, the MWIR fusion images were generated in three different sites. As a result, the boundaries of terrain and objects in the MWIR fusion images were emphasized to provide detailed thermal information of the interest area. Especially, the MWIR fusion image provided the thermal information of objects such as airplanes and ships which are hard to detect in the original MWIR images. This study demonstrated that the proposed method could generate a single image that combines visible details from an EO image and thermal information from an MWIR image, which contributes to increasing the usage of MWIR imagery.

• Korean Journal of Remote Sensing
• /
• v.39 no.5_3
• /
• pp.949-965
• /
• 2023

In this study, we propose a method to monitor the surface area of agricultural reservoirs in South Korea using Sentinel-1 synthetic aperture radar images and the deep learning model, Swin Transformer. Utilizing the Google Earth Engine platform, datasets from 2017 to 2021 were constructed for seven agricultural reservoirs, categorized into 700 K-ton, 900 K-ton, and 1.5 M-ton capacities. For four of the reservoirs, a total of 1,283 images were used for model training through shuffling and 5-fold cross-validation techniques. Upon evaluation, the Swin Transformer Large model, configured with a window size of 12, demonstrated superior semantic segmentation performance, showing an average accuracy of 99.54% and a mean intersection over union (mIoU) of 95.15% for all folds. When the best-performing model was applied to the datasets of the remaining three reservoirsfor validation, it achieved an accuracy of over 99% and mIoU of over 94% for all reservoirs. These results indicate that the Swin Transformer model can effectively monitor the surface area of agricultural reservoirs in South Korea.