• 한국정보전자통신기술학회논문지
• /
• 제12권5호
• /
• pp.467-474
• /
• 2019

본 이 논문에서는 NVIDIA Jetson TX1에서 YOLO v2 모델의 정확도를 유지하면서 FPS를 개선하는 방법을 제안한다. 일반적으로, 딥러닝 모델에서는 연산량을 줄여 처리 속도를 높이기 위해 파라미터들을 실수형에서 정수형으로 변환하여 정수 연산을 통해 속도를 높이거나 네트워크의 깊이를 감소시키는 방법을 사용한다. 그러나 이 방법들은 인식 정확도가 떨어질 수 있다. 이 논문에서는 YOLO v2 모델을 이용해 표정인식기를 개발하고 정확도 유지 시키기 위해 정수 연산이나 네트워크 깊이 감소를 사용하는 대신, 다음 세 가지 방법을 통해 연산량 및 메모리 소모를 줄인다. 첫 번째, $3{\times}3$ 필터를 $1{\times}1$ 필터로 교체하여 각 Layer 당 매개 변수 수를 9 분의 1로 줄인다. 두 번째, TensorRT의 추론 가속 기능 중 CBR (Convolution-Add Bias-Relu)을 통해 연산량을 줄이고, 마지막으로 TensorRT를 사용하여 반복되는 동일한 연산구조를 가진 레이어를 통합하여 메모리 소비를 줄인다. 시뮬레이션 결과, 기존 YOLO v2 모델에 비해 정확도는 1 % 감소했지만 FPS는 기존 3.9 FPS에서 11 FPS로 282%의 속도 향상을 보였다.

• 한국지진공학회논문집
• /
• 제26권5호
• /
• pp.191-202
• /
• 2022

Markov envelope as a theoretical solution of the parabolic wave equation with Markov approximation for the von Kármán type random medium is studied and approximated with the convolution of two probability density functions (pdf) of normal and gamma distributions considering the previous studies on the applications of Radiative Transfer Theory (RTT) and the analysis results of earthquake records. Through the approximation with gamma pdf, the constant shape parameter of 2 was determined regardless of the source distance r_o. This finding means that the scattering process has the property of an inhomogeneous single-scattering Poisson process, unlike the previous studies, which resulted in a homogeneous multiple-scattering Poisson process. Approximated Markov envelope can be treated as the normalized mean square (MS) envelope for ground acceleration because of the flat source Fourier spectrum. Based on such characteristics, the path duration is estimated from the approximated MS envelope and compared to the empirical formula derived by Boore and Thompson. The results clearly show that the path duration increases proportionately to r_o^1/2-r_o², and the peak value of the RMS envelope is attenuated by exp (-0.0033r_o), excluding the geometrical attenuation. The attenuation slope for r_o≤100 km is quite similar to that of effective attenuation for shallow crustal earthquakes, and it may be difficult to distinguish the contribution of intrinsic attenuation from effective attenuation. Slowly varying dispersive delay, also called the medium effect, represented by regular pdf, governs the path duration for the source distance shorter than 100 km. Moreover, the diffraction term, also called the distance effect because of scattering, fully controls the path duration beyond the source distance of 300 km and has a steep gradient compared to the medium effect. Source distance 100-300 km is a transition range of the path duration governing effect from random medium to distance. This means that the scattering may not be the prime cause of peak attenuation and envelope broadening for the source distance of less than 200 km. Furthermore, it is also shown that normal distribution is appropriate for the probability distribution of phase difference, as asserted in the previous studies.