• IEIE Transactions on Smart Processing and Computing
• /
• v.3 no.6
• /
• pp.366-371
• /
• 2014

This paper reports a machine learning approach for image object detection. Object detection and localization in a wild image, such as a STL-10 image dataset, is very difficult to implement using the traditional computer vision method. A convolutional neural network is a good approach for such wild image object detection. This paper presents an object detection application using a convolutional neural network with pretrained feature vector. This is a very simple and well organized hierarchical object abstraction model.

• ETRI Journal
• /
• v.42 no.4
• /
• pp.505-517
• /
• 2020

The increasing size and complexity of deep neural networks (DNNs) necessitate the development of efficient high-performance accelerators. An efficient memory structure and operating scheme provide an intuitive solution for high-performance accelerators along with dataflow control. Furthermore, the processing of various neural networks (NNs) requires a flexible memory architecture, programmable control scheme, and automated optimizations. We first propose an efficient architecture with flexibility while operating at a high frequency despite the large memory and PE-array sizes. We then improve the efficiency and usability of our architecture by automating the optimization algorithm. The experimental results show that the architecture increases the data reuse; a diagonal write path improves the performance by 1.44× on average across a wide range of NNs. The automated optimizations significantly enhance the performance from 3.8× to 14.79× and further provide usability. Therefore, automating the optimization as well as designing an efficient architecture is critical to realizing high-performance DNN accelerators.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.98-101
• /
• 2019

At present, deep convolutional neural networks have made a very important contribution in single-image super-resolution. Through the learning of the neural networks, the features of input images are transformed and combined to establish a nonlinear mapping of low-resolution images to high-resolution images. Some previous methods are difficult to train and take up a lot of memory. In this paper, we proposed a simple and compact deep recursive residual network learning the features for single image super resolution. Global residual learning and local residual learning are used to reduce the problems of training deep neural networks. And the recursive structure controls the number of parameters to save memory. Experimental results show that the proposed method improved image qualities that occur in previous methods.

• Journal of Korea Multimedia Society
• /
• v.20 no.5
• /
• pp.769-781
• /
• 2017

Deep learning networks like Convolutional Neural Networks (CNNs) show successful performances in many computer vision applications such as image classification, object detection, and so on. For implementation of deep learning networks in embedded system with limited processing power and memory, deep learning network may need to be simplified. However, simplified deep learning network cannot learn every possible scene. One realistic strategy for embedded deep learning network is to construct a simplified deep learning network model optimized for the scene images of the installation place. Then, automatic training will be necessitated for commercialization. In this paper, as an intermediate step toward automatic training under fisheye camera environments, we study more precise human localization in fisheye images, and propose an accurate human localization method, Automatic Ground-Truth Labelling Method (AGTLM). AGTLM first localizes candidate human object bounding boxes by utilizing GoogLeNet-LSTM approach, and after reassurance process by GoogLeNet-based CNN network, finally refines them more correctly and precisely(tightly) by applying saliency object detection technique. The performance improvement of the proposed human localization method, AGTLM with respect to accuracy and tightness is shown through several experiments.

• Journal of Positioning, Navigation, and Timing
• /
• v.10 no.1
• /
• pp.49-54
• /
• 2021

Indoor positioning system becomes of increasing interests due to the demands for accurate indoor location information where Global Navigation Satellite System signal does not approach. Wi-Fi access points (APs) built in many construction in advance helps developing a Wi-Fi Received Signal Strength Indicator (RSSI) based indoor localization. This localization method first collects pairs of position and RSSI measurement set, which is called fingerprint database, and then estimates a user's position when given a query measurement set by comparing the fingerprint database. The challenge arises from nonlinearity and noise on Wi-Fi RSSI measurements and complexity of handling a large amount of the fingerprint data. In this paper, machine learning techniques have been applied to implement Wi-Fi based localization. However, most of existing indoor localizations focus on single position estimation. The main contribution of this paper is to develop multi-target localization by using deep neural, which is beneficial when a massive crowd requests positioning service. This paper evaluates the proposed multilocalization based on deep learning from a multi-story building, and analyses its learning effect as increasing number of target positions.

• Journal of Advanced Navigation Technology
• /
• v.21 no.5
• /
• pp.435-442
• /
• 2017

The runway visual range affected by fog and so on is one of the important indicators to determine whether aircraft can take off and land at the airport or not. In the case of airports where transportation airplanes are operated, major weather forecasts including the runway visual range for local area have been released and provided to aviation workers for recognizing that. This paper proposes a runway visual range estimation model with a deep neural network applied recently to various fields such as image processing, speech recognition, natural language processing, etc. It is developed and implemented for estimating a runway visual range of local airport with a deep neural network. It utilizes the past actual weather observation data of the applied airfield for constituting the learning of the neural network. It can show comparatively the accurate estimation result when it compares the results with the existing observation data. The proposed model can be used to generate weather information on the airfield for which no other forecasting function is available.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.4
• /
• pp.305-312
• /
• 2023

Physics-Informed Neural Network (PINN) is used to invert bubble size distributions from attenuation losses. By considering a linear system for the bubble population inversion, Adaptive Learned Iterative Shrinkage Thresholding Algorithm (Ada-LISTA), which has been solved linear systems in image processing, is used as a neural network architecture in PINN. Furthermore, a regularization based on the linear system is added to a loss function of PINN and it makes a PINN have better generalization by a solution satisfying the bubble physics. To evaluate an uncertainty of bubble estimation, deep ensemble is adopted. 20 Ada-LISTAs with different initial values are trained using the same training dataset. During test with attenuation losses different from those in the training dataset, the bubble size distribution and corresponding uncertainty are indicated by average and variance of 20 estimations, respectively. Deep ensemble Ada-LISTA demonstrate superior performance in inverting bubble size distributions than the conventional convex optimization solver of CVX.

• Transactions of Materials Processing
• /
• v.29 no.5
• /
• pp.272-281
• /
• 2020

As the thickness of mobile communication devices is getting thinner, the size of the internal parts is also getting smaller. Among them, the battery case requires a high-level deep drawing technique because it has a rectangular shape with a large aspect ratio. In this study, the initial blank shape was optimized to minimize earing in a multi-stage deep drawing process using an artificial neural network(ANN). There has been no reported case of applying artificial neural network technology to the initial blank optimal design for a square case with large aspect ratio. The training data for ANN were obtained though simulation, and the model reliability was verified by performing comparative study with regression model using random sample test and goodness-of-fit test. Finally, the optimal design of the initial blank shape was performed through the verified ANN model.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2018.05a
• /
• pp.147-147
• /
• 2018

In this article, we use an open source software library: TensorFlow, developed for the purposes of conducting very complex machine learning and deep neural network applications. However, the system is general enough to be applicable in a wide variety of other domains as well. The proposed model based on a deep neural network model, LSTM (Long Short-Term Memory) to predict the river water level at Okcheon Station of the Guem River without utilization of rainfall - forecast information. For LSTM modeling, the input data is hourly water level data for 15 years from 2002 to 2016 at 4 stations includes 3 upstream stations (Sutong, Hotan, and Songcheon) and the forecasting-target station (Okcheon). The data are subdivided into three purposes: a training data set, a testing data set and a validation data set. The model was formulated to predict Okcheon Station water level for many cases from 3 hours to 12 hours of lead time. Although the model does not require many input data such as climate, geography, land-use for rainfall-runoff simulation, the prediction is very stable and reliable up to 9 hours of lead time with the Nash - Sutcliffe efficiency (NSE) is higher than 0.90 and the root mean square error (RMSE) is lower than 12cm. The result indicated that the method is able to produce the river water level time series and be applicable to the practical flood forecasting instead of hydrologic modeling approaches.

• Nuclear Engineering and Technology
• /
• v.51 no.2
• /
• pp.424-431
• /
• 2019

The aim of the present study was to develop model for detonation cell sizes prediction based on a deep artificial neural network of hydrogen, methane and propane mixtures with air and oxygen. The discussion about the currently available algorithms compared existing solutions and resulted in a conclusion that there is a need for a new model, free from uncertainty of the effective activation energy and the reaction length definitions. The model offers a better and more feasible alternative to the existing ones. Resulting predictions were validated against experimental data obtained during the investigation of detonation parameters, as well as with data collected from the literature. Additionally, separate models for individual mixtures were created and compared with the main model. The comparison showed no drawbacks caused by fitting one model to many mixtures. Moreover, it was demonstrated that the model may be easily extended by including more independent variables. As an example, dependency on pressure was examined. The preparation of experimental data for deep neural network training was described in detail to allow reproducing the results obtained and extending the model to different mixtures and initial conditions. The source code of ready to use models is also provided.