• ETRI Journal
• /
• v.44 no.6
• /
• pp.1034-1046
• /
• 2022

The learning-based multiview stereo (MVS) methods for three-dimensional (3D) reconstruction generally use 3D volumes for depth inference. The quality of the reconstructed depth maps and the corresponding point clouds is directly influenced by the spatial resolution of the 3D volume. Consequently, these methods produce point clouds with sparse local regions because of the lack of the memory required to encode a high volume of information. Here, we apply the atrous spatial pyramid pooling (ASPP) module in MVS methods to obtain dense feature maps with multiscale, long-range, contextual information using high receptive fields. For a given 3D volume with the same spatial resolution as that in the MVS methods, the dense feature maps from the ASPP module encoded with superior information can produce dense point clouds without a high memory footprint. Furthermore, we propose a 3D loss for training the MVS networks, which improves the predicted depth values by 24.44%. The ASPP module provides state-of-the-art qualitative results by constructing relatively dense point clouds, which improves the DTU MVS dataset benchmarks by 2.25% compared with those achieved in the previous MVS methods.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.161-164
• /
• 2022

Homomorphic Encryption (HE) schemes have been recently growing as a reliable solution to preserve users' information owe to maintaining and operating the user data in the encrypted state. In addition to that, several Neural Networks models merged with HE schemes have been developed as a prospective tool for privacy-preserving machine learning. Those mentioned works demonstrated that it is possible to match the accuracy of non-encrypted models but there is always a trade-off in the computation time. In this work, we evaluate the implementation of CKKS HE scheme operations over the layers of a LeNet5 convolutional inference model, however, owing to the limitations of the evaluation environment, the scope of this work is not to develop a complete LeNet5 encrypted model. The evaluation was performed using the MNIST dataset with Microsoft SEAL (MSEAL) open-source homomorphic encryption library ported version on Python (PyFhel). The behavior of the encrypted model, the limitations faced and a small description of related and future work is also provided.

• Advances in nano research
• /
• v.14 no.2
• /
• pp.141-154
• /
• 2023

Effects of viscoelastic foundation on vibration of curved-beam structure with clamped and simply-supported boundary conditions is investigated in this study. In doing so, a micro-scale laminate composite beam with two piezoelectric face layer with a carbon nanotube reinforces composite core is considered. The whole beam structure is laid on a viscoelastic substrate which normally occurred in actual conditions. Due to small scale of the structure non-classical elasticity theory provided more accurate results. Therefore, nonlocal strain gradient theory is employed here to capture both nano-scale effects on carbon nanotubes and microscale effects because of overall scale of the structure. Equivalent homogenous properties of the composite core is obtained using Halpin-Tsai equation. The equations of motion is derived considering energy terms of the beam and variational principle in minimizing total energy. The boundary condition is assumed to be clamped at one end and simply supported at the other end. Due to nonlinear terms in the equations of motion, semi-analytical method of general differential quadrature method is engaged to solve the equations. In addition, due to complexity in developing and solving equations of motion of arches, an artificial neural network is design and implemented to capture effects of different parameters on the inplane vibration of sandwich arches. At the end, effects of several parameters including nonlocal and gradient parameters, geometrical aspect ratios and substrate constants of the structure on the natural frequency and amplitude is derived. It is observed that increasing nonlocal and gradient parameters have contradictory effects of the amplitude and frequency of vibration of the laminate beam.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.18 no.2
• /
• pp.51-58
• /
• 2023

This paper proposes a deep learning method for estimating the heart rate from facial videos. Our proposed method estimates remote photoplethysmography (rPPG) signals to predict the heart rate. Although there have been proposed several methods for estimating rPPG signals, most previous methods can not be utilized in low-power single board computers due to their computational complexity. To address this problem, we construct a lightweight student model and employ a knowledge distillation technique to reduce the performance degradation of a deeper network model. The teacher model consists of 795k parameters, whereas the student model only contains 24k parameters, and therefore, the inference time was reduced with the factor of 10. By distilling the knowledge of the intermediate feature maps of the teacher model, we improved the accuracy of the student model for estimating the heart rate. Experiments were conducted on the UBFC-rPPG dataset to demonstrate the effectiveness of the proposed method. Moreover, we collected our own dataset to verify the accuracy and processing time of the proposed method on a real-world dataset. Experimental results on a NVIDIA Jetson Nano board demonstrate that our proposed method can infer the heart rate in real time with the mean absolute error of 2.5183 bpm.

• Structural Engineering and Mechanics
• /
• v.87 no.3
• /
• pp.231-242
• /
• 2023

A novel combinatorial type-2 adaptive neuro-fuzzy inference system (T2-ANFIS) and robust proportional integral derivative (PID) control framework for intelligent vibration mitigation of uncertain structural system is introduced. The fuzzy logic controllers (FLCs), are designed independently of the mathematical model of the system. The type-1 FLCs, have a limited ability to reduce the effect of uncertainty, due to their fuzzy sets with a crisp degree of membership. In real applications, the consequent part of the fuzzy rules is uncertain. The type-2 FLCs, are robust to the fuzzy rules and the process parameters due to the fuzzy degree of membership functions and footprint of uncertainty (FOU). The adaptivity of the proposed method is provided with the optimum tuning of the parameters using the neural network training algorithms. In our approach, the PID control force is obtained using the generalized type-2 neuro-fuzzy in such a way that the stability and robustness of the controller are guaranteed. The robust performance and stability of the presented framework are demonstrated in a numerical study for an eleven-story seismically-excited building structure combined with an active tuned mass damper (ATMD). The results indicate that the introduced type-2 neuro-fuzzy PID control scheme is effective to attenuate plant states in the presence of the structured and unstructured uncertainties, compared to the conventional, type-1 FLC, type-2 FLC, and type-1 neuro-fuzzy PID controllers.

• The Bulletin of The Korean Astronomical Society
• /
• v.46 no.1
• /
• pp.31.2-31.2
• /
• 2021

To estimate the contribution of quasars on keeping the IGM ionized, building a quasar luminosity function (LF) is necessary. Quasar LFs derived from multiple quasar surveys, however, are incompatible, especially for the faint regime, emphasizing the need for deep images. In this study, we construct quasar LF reaching M₁₄₅₀~-21.5 AB magnitude at z ~ 5, which is 1.5 mag deeper than previously reported LFs, using deep images from the Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP). We trained an artificial neural network (ANN) by inserting the colors as inputs to classify the quasars at z ~ 5 from the late-type stars and low-redshift galaxies. The accuracy of ANN is > 99 %. We also adopted the Bayesian information criterion to elaborate on the quasar-like objects. As a result, we recovered 5/5 confirmed quasars and remarkably minimized the contamination rate of high-redshift galaxies by up to six times compared to the selection using color selection alone. The constructed quasar parametric LF shows a flatter faint-end slope α=-127^+0.16_-0.15 similar to the recent LFs. The number of faint quasars (M₁₄₅₀ < -23.5) is too few to be the main contributor to IGM ionizing photons.

• Steel and Composite Structures
• /
• v.49 no.6
• /
• pp.645-666
• /
• 2023

Due to the steadily declining supply of natural coarse aggregates, the concrete industry has shifted to substituting coarse aggregates generated from byproducts and industrial waste. Oil palm shell is a substantial waste product created during the production of palm oil (OPS). When considering the usage of OPSC, building engineers must consider its uniaxial compressive strength (UCS). Obtaining UCS is expensive and time-consuming, machine learning may help. This research established five innovative hybrid AI algorithms to predict UCS. Aquila optimizer (AO) is used with methods to discover optimum model parameters. Considered models are artificial neural network (AO - ANN), adaptive neuro-fuzzy inference system (AO - ANFIS), support vector regression (AO - SVR), random forest (AO - RF), and extreme gradient boosting (AO - XGB). To achieve this goal, a dataset of OPS-produced concrete specimens was compiled. The outputs depict that all five developed models have justifiable accuracy in UCS estimation process, showing the remarkable correlation between measured and estimated UCS and models' usefulness. All in all, findings depict that the proposed AO - XGB model performed more suitable than others in predicting UCS of OPSC (with R², RMSE, MAE, VAF and A_15-index at 0.9678, 1.4595, 1.1527, 97.6469, and 0.9077). The proposed model could be utilized in construction engineering to ensure enough mechanical workability of lightweight concrete and permit its safe usage for construction aims.

• Journal of Korean Society of Transportation
• /
• v.19 no.5
• /
• pp.119-134
• /
• 2001

This paper proposed a dynamic transit vehicle simulation model and a dynamic transit passengers simulation model, which can simultaneously simulate the transit vehicles and passengers traveling on a transit network, and also developed an algorithm of dynamic departure time choice model based on individual passenger. The proposed model assumes that each passenger's behavior is heterogeneous based on stochastic process by relaxing the assumption of homogeneity among passengers and travelers have imperfect information and bounded rationality to more actually represent and to simulate each passenger's behavior. The proposed model integrated a inference and preference reforming procedure into the learning and decision making process in order to describe and to analyze the departure time choices of transit passengers. To analyze and evaluate the model an example transit line heading for work place was used. Numerical results indicated that in the model based on heterogeneous passengers the travelers' preference influenced more seriously on the departure time choice behavior, while in the model based on homogeneous passengers it does not. The results based on homogeneous passengers seemed to be unrealistic in the view of rational behavior. These results imply that the aggregated travel demand models such as the traditional network assignment models based on user equilibrium, assuming perfect information on the network, homogeneity and rationality, might be different from the real dynamic travel demand patterns occurred on actual network.

• Korean Journal of Cognitive Science
• /
• v.16 no.4
• /
• pp.369-386
• /
• 2005

Although Intelligent Tutoring System(ITS) offers individualized learning environment that overcome limited function of existent CAI, and consider many learners' variable, there is little development to be using at the sites of schools because of inefficiency of investment and absence of pedagogical content knowledge representation techniques. To solve these problem, we should study a method, which represents knowledge for ITS, and which reuses knowledge base. On the pedagogical content knowledge, the knowledge in education differs from knowledge in a general sense. In this paper, we shall primarily address the multi-complex structure of knowledge and explanation of learning vein using multi-complex structure. Multi-Complex, which is organized into nodes, clusters and uses by knowledge base. In addition, it grows a adaptive knowledge base by self-learning. Therefore, in this paper, we propose the 'Extended Neural Logic Network(X-Neuronet)', which is based on Neural Logic Network with logical inference and topological inflexibility in cognition structure, and includes pedagogical content knowledge and object-oriented conception, verify validity. X-Neuronet defines that a knowledge is directive combination with inertia and weights, and offers basic conceptions for expression, logic operator for operation and processing, node value and connection weight, propagation rule, learning algorithm.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.11
• /
• pp.1627-1634
• /
• 2021

This paper introduces a low-power compact ADC circuit for analog Convolutional filter for low-power neural network accelerator SOC. While convolutional neural network accelerators can speed up the learning and inference process, they have drawback of consuming excessive power and occupying large chip area due to large number of multiply-and-accumulate operators when implemented in complex digital circuits. To overcome these drawbacks, we implemented an analog convolutional filter that consists of an analog multiply-and-accumulate arithmetic circuit along with an ADC. This paper is focused on the design optimization of a low-power 8bit SAR ADC for the analog convolutional filter accelerator We demonstrate how to minimize the capacitor-array DAC, an important component of SAR ADC, which is three times smaller than the conventional circuit. The proposed ADC has been fabricated in CMOS 65nm process. It achieves an overall size of 1355.7㎛², power consumption of 2.6㎼ at a frequency of 100MHz, SNDR of 44.19 dB, and ENOB of 7.04bit.