• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• fall
• /
• pp.154-156
• /
• 2021

Depth-wise separable convolution 은 컴퓨터 자원이 제한된 환경에서 기존의 standard convolution을 대체하는데 강력하고, 효과적인 대안으로 잘 알려져 있다.[1] MobileNetV2 에서는 Inverted residual block을 소개한다. 이는 depth-wise separable convolution으로 인해 생기는 손실, 즉 channel 간의 데이터를 조합해 새로운 feature를 만들어낼 기회를 잃어버릴 때, 이를 depth-wise separable convolution 양단에 point-wise convolution(1×1 convolution)을 사용함으로써 극복해낸 block이다.[1] 하지만 1×1 convolution은 채널 수에 의존적(dependent)인 특징을 갖고 있고, 따라서 결국 네트워크가 깊어지면 깊어질수록 효율적이고(efficient) 가벼운(light weight) 네트워크를 만드는데 병목 현상(bottleneck)을 일으키고 만다. 이 논문에서는 channel-wise squeeze and excitation block(CSE)을 통해 1×1 convolution을 부분적으로 대체하는 방법을 통해 이 병목 현상을 해결한다.

• MALSORI
• /
• no.51
• /
• pp.85-98
• /
• 2004

It is almost impossible to generate the sound field effect in real time with the time-domain linear convolution because of its large multiplication operation requirement. To solve this, three methods are introduced to reduce the number of multiplication operations in this paper. Firstly, the time-domain linear convolution is replaced with the frequency-domain circular convolution. In other words, the linear convolution result can be derived from that of the circular convolution. This technique reduces the number of multiplication operations remarkably, Secondly, a subframe concept is introduced, i.e., one original frame is divided into several subframes. Then the FFT is executed for each subframe and, as a result, the number of multiplication operations can be reduced. Finally, the QFT is used in stead of the FFT. By combining all the above three methods into our final the SFE generation algorithm, the number of computations are reduced sufficiently and the real-time SFE generation becomes possible with a general PC.

• Journal of Broadcast Engineering
• /
• v.14 no.5
• /
• pp.543-550
• /
• 2009

The most widely used block convolution method is the overlap save algorithm (OSA), where a block of M data to be convolved with a filter is concatenated with the previous block and 2M-point FFT and multiplications are performed for this overlapped block. By discarding half of the results, we obtain linear convolution results from the circular convolution. This paper proposes a new transform which reduces the block size to only M for the block convolution. The proposed transform can be implemented as the M multiplications followed by M-point FFT Hence, existing efficient FFT libraries and hardware can be exploited for the implementation of proposed method. Since the required transform size is half that of the conventional method, the overall computational complexity is reduced. Also the reduced transform size results in the reduction of data access time and cash miss-hit ratio, and thus the overall CPU time is reduced. Experiments show that the proposed method requires less computation time than the conventional OSA.

• 제어로봇시스템학회:학술대회논문집
• /
• 2000.10a
• /
• pp.240-240
• /
• 2000

This paper presents two algorithms based on the concept of the frequency domain adaptive filter(FDAF). First the frequency domain recursive least squares(FRLS) algorithm with the overlap-save filtering technique is introduced. This minimizes the sum of exponentially weighted square errors in the frequency domain. To eliminate discrepancies between the linear convolution and the circular convolution, the overlap-save method is utilized. Second, the sliding method of data blocks is studied Co overcome processing delays and complexity roads of the FRLS algorithm. The size of the extended data block is twice as long as the filter tap length. It is possible to slide the data block variously by the adjustable hopping index. By selecting the hopping index appropriately, we can take a trade-off between the convergence rate and the computational complexity. When the input signal is highly correlated and the length of the target FIR filter is huge, the FRLS algorithm based on the block processing technique has good performances in the convergence rate and the computational complexity.

• Journal of Korea Multimedia Society
• /
• v.23 no.7
• /
• pp.803-811
• /
• 2020

This paper proposes an image classification algorithm that transforms the number of convolution layers in the residual block of ResNet, CNN's representative method. The proposed method modified the structure of 34/50 layer of ResNet structure. First, we analyzed the performance of small and many convolution layers for the structure consisting of only shortcut and 3 × 3 convolution layers for 34 and 50 layers. And then the performance was analyzed in the case of small and many cases of convolutional layers for the bottleneck structure of 50 layers. By applying the results, the best classification method in the residual block was applied to construct a 34-layer simple structure and a 50-layer bottleneck image classification model. To evaluate the performance of the proposed image classification model, the results were analyzed by applying to the cifar10 dataset. The proposed 34-layer simple structure and 50-layer bottleneck showed improved performance over the ResNet-110 and Densnet-40 models.

• ETRI Journal
• /
• v.4 no.3
• /
• pp.18-29
• /
• 1982

In a block adaptive filtering procedure, the filter coefficients are adjusted once per each output block while maintaining performance comparable to that of widely used LMS adaptive filtering in which the filter coefficients are adjusted once per each output data sample. An efficient implementation of block adaptive filter is possible by means of discrete transform technique which has cyclic convolution property and fast algorithms. In this paper, the block adaptive filtering using Fermat Number Transform (FNT) is investigated to exploit the computational efficiency and less quantization effect on the performance compared with finite precision FFT realization. And this has been verified by computer simulation for several applications including adaptive channel equalizer and system identification.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.508-510
• /
• 2019

본 연구에서는 딥 러닝을 이용하여 완전 자동화된 다중 클래스 전체 심장 분할 알고리즘을 제안하였다. 제안된 방법은 recurrent convolutional block과 residual multi-dilated block을 삽입하여 기존 U-Net을 개선한 인공신경망 모델을 사용하였다. 평가는 자동화 분석 결과와 수동 평가를 비교하였다. 그 결과 96.88%의 평균 DSC, 95.60%의 정확도, 97.00%의 recall을 얻었다. 이 실험 결과는 제안된 방법이 다양한 심장 구조에서 효과적으로 구분되어 수행되었음을 알 수 있다. 본 연구에서 제안된 알고리즘이 의사와 방사선 의사가 영상을 판독하거나 임상 결정을 내리는데 보조적 역할을 할 것을 기대한다.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.14 no.1
• /
• pp.1-9
• /
• 2019

In this paper, we propose a method of generating a 2D gray image from LiDAR 3D reflection intensity. The proposed method uses the Fully Convolutional Network (FCN) to generate the gray image from 2D reflection intensity which is projected from LiDAR 3D intensity. Both encoder and decoder of FCN are configured with several convolution blocks in the symmetric fashion. Each convolution block consists of a convolution layer with $3{\times}3$ filter, batch normalization layer and activation function. The performance of the proposed method architecture is empirically evaluated by varying depths of convolution blocks. The well-known KITTI data set for various scenarios is used for training and performance evaluation. The simulation results show that the proposed method produces the improvements of 8.56 dB in peak signal-to-noise ratio and 0.33 in structural similarity index measure compared with conventional interpolation methods such as inverse distance weighted and nearest neighbor. The proposed method can be possibly used as an assistance tool in the night-time driving system for autonomous vehicles.

• Journal of Advanced Navigation Technology
• /
• v.15 no.6
• /
• pp.1068-1074
• /
• 2011

In this paper, the proposed scheme is STBC(space-time block codes) system combined with convolution code which is the most popular channel coding to ensure the reliability of data transmission for a high data rate wireless communication. The STBC is one of MIMO(multi-input multi-output) techniques. In addition, this scheme uses a modified viterbi algorithm in order to get a high system gain when data is transmitted. Because we combine STBC and convolution code, the proposed scheme has a little high quantity of computation but it can get a maximal diversity gain of STBC and a high coding gain of convolution code at the same time. Unlike existing viterbi docoding algorithm using Hamming distance in order to calculate branch matrix, the modified viterbi algorithm uses Euclidean distance value between received symbol and reference symbol. Simulation results show that the modified viterbi algorithm improved gain 7.5 dB on STBC 2Tx-2Rx at $BER=10^{-2}$. Therefore the proposed scheme using STBC combined with convolution code can improve the transmission reliability and transmission efficiency.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2019.07a
• /
• pp.281-284
• /
• 2019

본 논문은 CNN (Convolution Neural Networks)의 첫 번째 컨볼루션층(convolution layer)을 RGB-csb(RGB channel separation block)로 대체하여 입력 영상의 RGB 값을 특징 맵에 적용시켜 정확성을 제고시킬 수 있는 선행연구 결과에 추가적으로, 훈련 및 시험 영상 수에 따른 분석을 통하여 정확도 향상방법을 제안한다. 제안한 방법은 영상의 개수가 작을수록 각 학습 간의 정확도 편차가 크게 나타나는 불안정성은 있지만 기존 CNN모델에 비하여 정확도 차이가 증가함을 알 수 있다.