• 반도체공학회 논문지
• /
• 제2권1호
• /
• pp.1-8
• /
• 2024

깊이별 분리 합성곱 (Depthwise Separable Convolution)을 처리할 때, processing element (PE)의 저활용성은 시스톨릭 어레이 (SA)의 한계점 중 하나이다. 본 연구에서는 깊이별 합성곱의 처리량을 극대화하기 위한 새로운 SA 아키텍처를 제안한다. 더불어, 제안된 SA 는 깊이별 합성곱 계산 중에 유휴 PE 에서 후속 점별 합성곱 (pointwise convolution)을 수행하여 활용도를 증가시킨다. 모든 깊이별 합성곱 연산 후에는 모든 PE 를 활용하여 나머지 점별 합성곱 연산의 속도를 향상시킨다. 결과적으로, 제안된 128×128 SA 는 MobileNetV3 연산 시, 기본 SA 및 RiSA 와 비교하여 속도가 4.05 배, 1.75 배 향상되고, 에너지 소비량을 각각 66.7 %, 25.4 % 감소한다.

• International Journal of Computer Science & Network Security
• /
• 제21권11호
• /
• pp.17-22
• /
• 2021

Recently, image super-resolution techniques used in convolutional neural networks (CNN) have led to remarkable performance in the research area of digital image processing applications and computer vision tasks. Convolutional layers stacked on top of each other can design a more complex network architecture, but they also use more memory in terms of the number of parameters and introduce the vanishing gradient problem during training. Furthermore, earlier approaches of single image super-resolution used interpolation technique as a pre-processing stage to upscale the low-resolution image into HR image. The design of these approaches is simple, but not effective and insert the newer unwanted pixels (noises) in the reconstructed HR image. In this paper, authors are propose a novel single image super-resolution architecture based on synchronized depthwise separable convolution with Dense Skip Connection Block (DSCB). In addition, unlike existing SR methods that only rely on single path, but our proposed method used the synchronizes path for generating the SISR image. Extensive quantitative and qualitative experiments show that our method (SDCN) achieves promising improvements than other state-of-the-art methods.

• 전기전자학회논문지
• /
• 제28권3호
• /
• pp.375-381
• /
• 2024

본 논문에서는 배달 서비스 로봇 응용을 위한 LiDAR 센서 기반 경량화된 객체 분류 시스템을 제안한다. 3차원 포인트 클라우드 데이터를 Pillar Feature Network (PFN)을 사용하여 2차원 pseudo image로 인코딩한 후, Depthwise Separable Convolution Neural Network (DS-CNN)에 기반하여 설계된 네트워크를 통해 객체 분류를 수행하는 경량화된 시스템을 설계하였다. 구현 결과, 설계한 분류 네트워크의 파라미터 수와 Multiply-Accumulate (MAC) 연산 수는 각각 9.08K 및 3.49M이며, 94.94%의 분류 정확도를 지원 가능함을 확인하였다.

• 응용통계연구
• /
• 제34권6호
• /
• pp.979-993
• /
• 2021

음향 장면 분류는 오디오 파일이 녹음된 환경이 어디인지 분류하는 문제이다. 이는 음향 장면 분류와 관련한 대회인 DCASE 대회에서 꾸준하게 연구되었던 분야이다. 실제 응용 분야에 음향 장면 분류 문제를 적용할 때, 모델의 복잡도를 고려하여야 한다. 특히 경량 기기에 적용하기 위해서는 경량 딥러닝 모델이 필요하다. 우리는 경량 기술이 적용된 여러 모델을 비교하였다. 먼저 log mel-spectrogram, deltas, delta-deltas 피쳐를 사용한 합성곱 신경망(CNN) 기반의 기본 모델을 제안하였다. 그리고 원래의 합성곱 층을 depthwise separable convolution block, linear bottleneck inverted residual block과 같은 효율적인 합성곱 블록으로 대체하고, 각 모델에 대하여 Quantization를 적용하여 경량 모델을 제안하였다. 경량화 기술을 고려한 모델은 기본 모델에 대비하여 성능이 비슷하거나 조금 낮은 성능을 보였지만, 모델 사이즈는 503KB에서 42.76KB로 작아진 것을 확인하였다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제17권3호
• /
• pp.861-880
• /
• 2023

Semantic segmentation of road scene is the key technology of autonomous driving, and the improvement of convolutional neural network architecture promotes the improvement of model segmentation performance. The existing convolutional neural network has the simplification of learning knowledge and the complexity of the model. To address this issue, we proposed a road scene semantic segmentation algorithm based on multi-task collaborative learning. Firstly, a depthwise separable convolution atrous spatial pyramid pooling is proposed to reduce model complexity. Secondly, a collaborative learning framework is proposed involved with saliency detection, and the joint loss function is defined using homoscedastic uncertainty to meet the new learning model. Experiments are conducted on the road and nature scenes datasets. The proposed method achieves 70.94% and 64.90% mIoU on Cityscapes and PASCAL VOC 2012 datasets, respectively. Qualitatively, Compared to methods with excellent performance, the method proposed in this paper has significant advantages in the segmentation of fine targets and boundaries.

• ETRI Journal
• /
• 제45권4호
• /
• pp.666-677
• /
• 2023

In the last few years, convolutional neural networks (CNNs) have demonstrated good performance while solving various computer vision problems. However, since CNNs exhibit high computational complexity, signal processing is performed on the server side. To reduce the computational complexity of CNNs for edge computing, a lightweight algorithm, such as a MobileNet, is proposed. Although MobileNet is lighter than other CNN models, it commonly achieves lower classification accuracy. Hence, to find a balance between complexity and accuracy, additional hyperparameters for adjusting the size of the model have recently been proposed. However, significantly increasing the number of parameters makes models dense and unsuitable for devices with limited computational resources. In this study, we propose a novel MobileNet architecture, in which the number of parameters is adaptively increased according to the importance of feature maps. We show that our proposed network achieves better classification accuracy with fewer parameters than the conventional MobileNet.

• 사물인터넷융복합논문지
• /
• 제10권4호
• /
• pp.43-53
• /
• 2024

토마토에는 리코펜, β-카로틴 및 비타민 C와 같은 영양소가 풍부하고 세계적으로 많이 소비되는 채소 중 하나이다. 그러나 종종 생물학적 및 환경적 스트레스 요인으로 인해 수확량 손실이 발생한다. 전통적인 작물 건강 평가는 오류가 발생하기 쉽고 대규모 생산에 비효율적이다. 이러한 문제를 해결하기 위해 건강 상태에 대해 1~5로 주석을 메긴 토마토 전체 생육기간을 다루는 포괄적인 데이터 세트를 수집하였다. 우리는 Channel-wise attention과 Grouped convolution을 사용한 Attention-Enhanced DS-ResNet 아키텍처와 새로운 학습 기법을 제안한다. 우리의 모델은 5-fold 교차 검증을 사용하여 전체 정확도 80.2%를 달성하여 작물의 건강 상태를 정확하게 분류하는데 있어 견고성을 보여주었다.

• Smart Structures and Systems
• /
• 제32권3호
• /
• pp.179-193
• /
• 2023

This study explores an alternative to the existing centralized process for data anomaly detection in modern Internet of Things (IoT)-based structural health monitoring (SHM) systems. An edge intelligence framework is proposed for the early detection and classification of various data anomalies facilitating quality enhancement of acquired data before transmitting to a central system. State-of-the-art deep neural network pruning techniques are investigated and compared aiming to significantly reduce the network size so that it can run efficiently on resource-constrained edge devices such as wireless smart sensors. Further, depthwise separable convolution (DSC) is invoked, the integration of which with advanced structural pruning methods exhibited superior compression capability. Last but not least, quantization-aware training (QAT) is adopted for faster processing and lower memory and power consumption. The proposed edge intelligence framework will eventually lead to reduced network overload and latency. This will enable intelligent self-adaptation strategies to be employed to timely deal with a faulty sensor, minimizing the wasteful use of power, memory, and other resources in wireless smart sensors, increasing efficiency, and reducing maintenance costs for modern smart SHM systems. This study presents a theoretical foundation for the proposed framework, the validation of which through actual field trials is a scope for future work.

• Journal of Information Processing Systems
• /
• 제19권5호
• /
• pp.614-630
• /
• 2023

Dead trees significantly impact forest production and the ecological environment and pose constraints to the sustainable development of forests. A lightweight YOLOv4 dead tree detection algorithm based on unmanned aerial vehicle images is proposed to address current limitations in dead tree detection that rely mainly on inefficient, unsafe and easy-to-miss manual inspections. An improved logarithmic transformation method was developed in data pre-processing to display tree features in the shadows. For the model structure, the original CSPDarkNet-53 backbone feature extraction network was replaced by MobileNetV3. Some of the standard convolutional blocks in the original extraction network were replaced by depthwise separable convolution blocks. The new ReLU6 activation function replaced the original LeakyReLU activation function to make the network more robust for low-precision computations. The K-means++ clustering method was also integrated to generate anchor boxes that are more suitable for the dataset. The experimental results show that the improved algorithm achieved an accuracy of 97.33%, higher than other methods. The detection speed of the proposed approach is higher than that of YOLOv4, improving the efficiency and accuracy of the detection process.