• International journal of advanced smart convergence
• /
• 제7권2호
• /
• pp.95-100
• /
• 2018

In this paper, we propose a binarized multi-scale module to accelerate the speed of the pose estimating deep neural network. Recently, deep learning is also used for fine-tuned tasks such as pose estimation. One of the best performing pose estimation methods is based on the usage of two neural networks where one computes the heat maps of the body parts and the other computes the part affinity fields between the body parts. However, the convolution filtering with a large kernel filter takes much time in this model. To accelerate the speed in this model, we propose to change the large kernel filters with binarized multi-scale modules. The large receptive field is captured by the multi-scale structure which also prevents the dropdown of the accuracy in the binarized module. The computation cost and number of parameters becomes small which results in increased speed performance.

• 방송공학회논문지
• /
• 제23권3호
• /
• pp.383-394
• /
• 2018

본 논문에서는 최근 활발히 연구되고 있는 딥 러닝 기반의 이미지와 비디오 압축 기술에 대해 살펴본다. 딥 러닝 기반의 이미지 압축 기술은 심층 신경망에 압축 대상 이미지를 입력하고 반복적 또는 일괄적 방식으로 은닉 벡터를 추출하여 부호화한다. 이미지 압축 효율을 높이기 위해 심층 신경망은 복원 이미지의 화질은 높이면서 부호화된 은닉 벡터가 보다 적은 비트로 표현될 수 있도록 학습된다. 이러한 기술들은 특히 저 비트율에서 기존의 이미지 압축 기술에 비해 뛰어난 화질의 이미지를 생성할 수 있다. 한편, 딥 러닝 기반의 비디오 압축 기술은 압축 대상 비디오를 직접 입력하여 처리하기 보다는 기존 비디오 코덱의 압축 툴 성능을 개선하는 접근법을 취하고 있다. 본 논문에서 소개하는 심층 신경망 기술들은 최신 비디오 코덱의 인루프 필터를 대체하거나 추가적인 후처리 필터로 사용되어 복원 영상의 화질 개선을 통해 압축 효율을 향상시킨다. 마찬가지로, 화면 내 예측 및 부호화에 적용된 심층 신경망 기술들은 기존 화면 내 예측 툴과 함께 사용되어 예측 정확도를 높이거나 새로운 화면 내 부호화 과정을 추가함으로써 압축 효율을 향상 시킨다.

• International Journal of Internet, Broadcasting and Communication
• /
• 제13권1호
• /
• pp.161-167
• /
• 2021

This paper proposes a new personalized HRTF estimation method which is based on a deep neural network (DNN) model and improved elevation reproduction using a notch filter. In the previous study, a DNN model was proposed that estimates the magnitude of HRTF by using anthropometric measurements [1]. However, since this method uses zero-phase without estimating the phase, it causes the internalization (i.e., the inside-the-head localization) of sound when listening the spatial sound. We devise a method to estimate both the magnitude and phase of HRTF based on the DNN model. Personalized HRIR was estimated using the anthropometric measurements including detailed data of the head, torso, shoulders and ears as inputs for the DNN model. After that, the estimated HRIR was filtered with an appropriate notch filter to improve elevation reproduction. In order to evaluate the performance, both of the objective and subjective evaluations are conducted. For the objective evaluation, the root mean square error (RMSE) and the log spectral distance (LSD) between the reference HRTF and the estimated HRTF are measured. For subjective evaluation, the MUSHRA test and preference test are conducted. As a result, the proposed method can make listeners experience more immersive audio than the previous methods.

• 반도체디스플레이기술학회지
• /
• 제19권3호
• /
• pp.12-18
• /
• 2020

In semiconductor manufacturing, defect detection is critical to maintain high yield. Currently, computer vision systems used in semiconductor photo lithography still have adopt to digital image processing algorithm, which often occur inspection faults due to sensitivity to external environment. Thus, we intend to handle this problem by means of using Mask R-CNN instead of digital image processing algorithm. Additionally, Mask R-CNN can be trained with image dataset pre-processed by means of the specific designed digital image filter to extract the enhanced feature map of Convolutional Neural Network (CNN). Our approach converged advantage of digital image processing and instance segmentation with deep learning yields more efficient semiconductor photo lithography inspection system than conventional system.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제11권6호
• /
• pp.3029-3045
• /
• 2017

The main processes of a cognitive radio system include spectrum sensing, spectrum decision, spectrum sharing, and spectrum conversion. Experimental results show that these stages introduce a time delay that affects the spectrum sensing accuracy, reducing its efficiency. To reduce the time delay, the frequency spectrum prediction was proposed to alleviate the burden on the spectrum sensing. In this paper, the deep recurrent neural network (DRNN) was proposed to predict the spectrum of multiple time slots, since the existing methods only predict the spectrum of one time slot. The continuous state of a channel is divided into a many time slots, forming a time series of the channel state. Since there are more hidden layers in the DRNN than in the RNN, the DRNN has fading memory in its bottom layer as well as in the past input. In addition, the extended Kalman filter was used to train the DRNN, which overcomes the problem of slow convergence and the vanishing gradient of the gradient descent method. The spectrum prediction based on the DRNN was verified with a WiFi signal, and the error of the prediction was analyzed. The simulation results proved that the multiple slot spectrum prediction improved the spectrum efficiency and reduced the energy consumption of spectrum sensing.

• 로봇학회논문지
• /
• 제14권2호
• /
• pp.87-93
• /
• 2019

This paper presents a 6-DOF relocalization using a 3D laser scanner and a monocular camera. A relocalization problem in robotics is to estimate pose of sensor when a robot revisits the area. A deep convolutional neural network (CNN) is designed to regress 6-DOF sensor pose and trained using both RGB image and 3D point cloud information in end-to-end manner. We generate the new input that consists of RGB and range information. After training step, the relocalization system results in the pose of the sensor corresponding to each input when a new input is received. However, most of cases, mobile robot navigation system has successive sensor measurements. In order to improve the localization performance, the output of CNN is used for measurements of the particle filter that smooth the trajectory. We evaluate our relocalization method on real world datasets using a mobile robot platform.

• 디지털콘텐츠학회 논문지
• /
• 제18권1호
• /
• pp.183-190
• /
• 2017

본 논문에서는 깊은 신경망을 이용한 오디오 이벤트 검출 방법을 제안한다. 오디오 입력의 매 프레임에 대한 오디오 이벤트 확률을 feed-forward 신경망을 적용하여 생성한다. 매 프레임에 대하여 멜 스케일 필터 뱅크 특징을 추출한 후, 해당 프레임의 전후 프레임으로부터의 특징벡터들을 하나의 특징벡터로 결합하고 이를 feed-forward 신경망의 입력으로 사용한다. 깊은 신경망의 출력층은 입력 프레임 특징값에 대한 오디오 이벤트 확률값을 나타낸다. 연속된 5개 이상의 프레임에서의 이벤트 확률값이 임계값을 넘을 경우 해당 구간이 오디오 이벤트로 검출된다. 검출된 오디오 이벤트는 1초 이내에 동일 이벤트로 검출되는 동안 하나의 오디오 이벤트로 유지된다. 제안된 방법으로 구현된 오디오 이벤트 검출기는 UrbanSound8K와 BBC Sound FX자료에서의 20개 오디오 이벤트에 대하여 71.8%의 검출 정확도를 보였다.

• 한국지진공학회논문집
• /
• 제25권2호
• /
• pp.71-81
• /
• 2021

This paper presents a real-time, false-pick filter based on deep learning to reduce false alarms of an onsite Earthquake Early Warning (EEW) system. Most onsite EEW systems use P-wave to predict S-wave. Therefore, it is essential to properly distinguish P-waves from noises or other seismic phases to avoid false alarms. To reduce false-picks causing false alarms, this study made the EEWNet Part 1 'False-Pick Filter' model based on Convolutional Neural Network (CNN). Specifically, it modified the Pick_FP (Lomax et al.) to generate input data such as the amplitude, velocity, and displacement of three components from 2 seconds ahead and 2 seconds after the P-wave arrival following one-second time steps. This model extracts log-mel power spectrum features from this input data, then classifies P-waves and others using these features. The dataset consisted of 3,189,583 samples: 81,394 samples from event data (727 events in the Korean Peninsula, 103 teleseismic events, and 1,734 events in Taiwan) and 3,108,189 samples from continuous data (recorded by seismic stations in South Korea for 27 months from 2018 to 2020). This model was trained with 1,826,357 samples through balancing, then tested on continuous data samples of the year 2019, filtering more than 99% of strong false-picks that could trigger false alarms. This model was developed as a module for USGS Earthworm and is written in C language to operate with minimal computing resources.

• 농업과학연구
• /
• 제47권4호
• /
• pp.1109-1122
• /
• 2020

The aim of this study was to develop a marbling classification and prediction model using small parts of sirloin images based on a deep learning algorithm, namely, a convolutional neural network (CNN). Samples were purchased from a commercial slaughterhouse in Korea, images for each grade were acquired, and the total images (n = 500) were assigned according to their grade number: 1++, 1+, 1, and both 2 & 3. The image acquisition system consists of a DSLR camera with a polarization filter to remove diffusive reflectance and two light sources (55 W). To correct the distorted original images, a radial correction algorithm was implemented. Color images of sirloins of Hanwoo (mixed with feeder cattle, steer, and calf) were divided and sub-images with image sizes of 161 × 161 were made to train the marbling prediction model. In this study, the convolutional neural network (CNN) has four convolution layers and yields prediction results in accordance with marbling grades (1++, 1+, 1, and 2&3). Every single layer uses a rectified linear unit (ReLU) function as an activation function and max-pooling is used for extracting the edge between fat and muscle and reducing the variance of the data. Prediction accuracy was measured using an accuracy and kappa coefficient from a confusion matrix. We summed the prediction of sub-images and determined the total average prediction accuracy. Training accuracy was 100% and the test accuracy was 86%, indicating comparably good performance using the CNN. This study provides classification potential for predicting the marbling grade using color images and a convolutional neural network algorithm.

• 한국음향학회지
• /
• 제36권6호
• /
• pp.407-412
• /
• 2017

본 논문에서는 주목 메커니즘 기반의 심층 신경망을 사용한 음성 감정인식 방법을 제안한다. 제안하는 방식은 CNN(Convolution Neural Networks), GRU(Gated Recurrent Unit), DNN(Deep Neural Networks)의 결합으로 이루어진 심층 신경망 구조와 주목 메커니즘으로 구성된다. 음성의 스펙트로그램에는 감정에 따른 특징적인 패턴이 포함되어 있으므로 제안하는 방식에서는 일반적인 CNN에서 컨벌루션 필터를 tuned Gabor 필터로 사용하는 GCNN(Gabor CNN)을 사용하여 패턴을 효과적으로 모델링한다. 또한 CNN과 FC(Fully-Connected)레이어 기반의 주목 메커니즘을 적용하여 추출된 특징의 맥락 정보를 고려한 주목 가중치를 구해 감정인식에 사용한다. 본 논문에서 제안하는 방식의 검증을 위해 6가지 감정에 대해 인식 실험을 진행하였다. 실험 결과, 제안한 방식이 음성 감정인식에서 기존의 방식보다 더 높은 성능을 보였다.