Article
Engineering, Electrical & Electronic
Xinyu Wang, Liming Yuan, Haixia Xu, Xianbin Wen
Summary: This article proposes a channel-spatial attention mechanism based on a depthwise separable convolution (CSDS) network for aerial scene classification. Experimental results on three public datasets show that the CSDS network achieves comparable performance to other state-of-the-art methods. Visualization of feature extraction results and ablation experiments demonstrate the powerful feature learning and representation capabilities of the proposed CSDS network.
IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
(2021)
Article
Computer Science, Artificial Intelligence
Jun-Gi Jang, Chun Quan, Hyun Dong Lee, U. Kang
Summary: This paper proposes Falcon, an accurate and lightweight method to compress CNN based on depthwise separable convolution. Falcon interprets existing convolution methods based on depthwise separable convolution using a proposed mathematical formulation called generalized elementwise product (GEP). Experimental results show that Falcon achieves higher accuracy than existing methods while reducing the number of parameters and FLOPs of standard convolution.
KNOWLEDGE AND INFORMATION SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Xiaoyu Kong, Ke Zhang
Summary: Human behavior is influenced by emotions, and predicting behavior through emotion classification from text is significant for decision-making. Efficiently extracting emotional tendencies from text data is a challenge, but a upgraded CNN model proposed in this study improves the downsides and shows better performance in sentiment analysis tasks.
PEERJ COMPUTER SCIENCE
(2023)
Article
Engineering, Electrical & Electronic
Ming Gao, Pengjiang Qian
Summary: This study proposes a convolutional neural network algorithm based on a multilayer network model, which uses Exponential Linear Units-guided Depthwise Separable Convolution to extract spectral and spatial features from hyperspectral images. The feature maps are then fed into a cross-attention mechanism for weight allocation to optimize local feature information and computational resource allocation. Experimental results show that this network model achieves accurate hyperspectral image classification with improved computing efficiency.
Article
Engineering, Electrical & Electronic
Chenghong Xiao, Shuyuan Yang, Zhixi Feng
Summary: In this article, a novel end-to-end automatic modulation classification (AMC) model called complex-valued depthwise separable convolutional neural network (CDSCNN) is proposed, which adopts complex-valued operation units for tailored feature learning for AMC. With an overall accuracy of 62.63% on the RadioML2016.10a dataset, CDSCNN outperforms its counterparts by 1%-11%. After fine-tuning on the RadioML2016.10b dataset, the overall accuracy reaches 63.15%, demonstrating the robust recognition and generalization capability of CDSCNN. Moreover, CDSCNN exhibits lower model complexity compared to other methods.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2023)
Article
Automation & Control Systems
Fan-Hsun Tseng, Kuo-Hui Yeh, Fan-Yi Kao, Chi -Yuan Chen
Summary: This study proposes a new artificial intelligence model, MiniNet, which reduces computations and shortens training time under limited hardware resources. MiniNet utilizes depthwise and pointwise convolutions and incorporates dense connection technique and Squeeze-and-Excitation operations. Experimental results demonstrate that MiniNet significantly reduces parameters and training time while achieving high accuracy.
Article
Engineering, Biomedical
Yi Lu, Mingfeng Jiang, Liying Wei, Jucheng Zhang, Zhikang Wang, Bo Wei, Ling Xia
Summary: Arrhythmia is a major cause of morbidity and mortality among cardiac patients, and a depthwise separable convolutional neural network with focal loss method was proposed for automated arrhythmia classification with imbalance ECG dataset, showing improved performance and reduced parameters. The model achieved an overall macro average F1-score of 0.79 on the MIT-BIH arrhythmia database, representing an improvement in arrhythmia classification.
BIOMEDICAL SIGNAL PROCESSING AND CONTROL
(2021)
Article
Computer Science, Theory & Methods
Gangzhao Lu, Weizhe Zhang, Zheng Wang
Summary: This article optimizes the depthwise separable convolution for small-batch model training and model inference scenarios. Through the design of two new algorithms, the utilization and performance of the convolution operation are improved, and experimental results demonstrate the effectiveness of the approach on two GPU platforms.
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
(2022)
Article
Computer Science, Information Systems
Jiye Huang, Xin Liu, Tongdong Guo, Zhijin Zhao
Summary: A high-performance DSC hardware accelerator based on FPGAs is proposed in this paper, which improves the utilization of computational resources and reduces on-chip memory occupancy for DSC. Experimental results show that the proposed accelerator achieves a performance of 205.1 FPS, 128.8 GFLOPS, and 0.24 GOPS/DSP for input images of size 224x224x3.
Article
Engineering, Electrical & Electronic
Baoting Li, Hang Wang, Xuchong Zhang, Jie Ren, Longjun Liu, Hongbin Sun, Nanning Zheng
Summary: This paper introduces two efficient dynamic design techniques for improving the hardware efficiency and performance of DSC-based lightweight CNN accelerators, and their effectiveness and efficiency have been extensively evaluated. Experimental results demonstrate that the proposed architectural techniques can significantly reduce the on-chip buffer size and enhance the performance of convolution calculation.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS I-REGULAR PAPERS
(2021)
Article
Computer Science, Artificial Intelligence
Feng Liang, Zhichao Tian, Ming Dong, Shuting Cheng, Li Sun, Hai Li, Yiran Chen, Guohe Zhang
Summary: This paper proposes the use of LPPC convolution kernels to reduce computational complexities and storage costs of neural networks, designing four types of LPPC kernels, with experimental results showing that Type-I LPPC kernels can compress networks better with a slight reduction in accuracy.
Article
Computer Science, Artificial Intelligence
Jiaxing He, Xiaodan Wang, Yafei Song, Qian Xiang
Summary: To solve the issue of feature extraction in network intrusion detection caused by large-scale high-dimensional traffic data, we propose PyDSC-IDS, a method based on the variational Gaussian model and one-dimensional Pyramid Depthwise Separable Convolution neural network. PyDSC-IDS uses VGM and OneHot encode technologies to preprocess the original dataset and decompose complex features into simpler ones. The experimental results demonstrate the effectiveness of PyDSC-IDS in improving detection accuracy and reducing network complexity.
Article
Computer Science, Artificial Intelligence
Pravendra Singh, Pratik Mazumder, Vinay P. Namboodiri
Summary: Convolutional layers are essential in generating valuable output features and aiding deep learning methods in solving complex problems; Pointwise convolutions are used primarily for parameter reduction, but they overlook spatial information; The novel alternative design efficiently extracts and integrates spatial information, significantly improving network performance.
PATTERN RECOGNITION
(2022)
Article
Medicine, General & Internal
Shing-Yun Jung, Chia-Hung Liao, Yu-Sheng Wu, Shyan-Ming Yuan, Chuen-Tsai Sun
Summary: This research proposes a feature engineering process to extract dedicated features for the DS-CNN model to classify lung sounds accurately and efficiently. The fusion of STFT and MFCC features with DS-CNN leads to faster inference speed and higher accuracy on edge devices, indicating a potential model design for accurate AI-aided detection of lung diseases.
Article
Environmental Sciences
Yingpeng Dai, Chenglin Li, Xiaohang Su, Hongxian Liu, Jiehao Li
Summary: A unique convolution module, consisting of concatenation pointwise convolution and multi-scale depthwise convolution, is proposed for real-time semantic segmentation. This module strengthens the non-linear relationship between input and output and extracts multi-scale spatial features for describing objects.