Zhao Zhongkai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhao Zhongkai has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Tatsuhito Hasegawa
Graduate School of Engineering, University of Fukui, Fukui, Japan
2 shared publications - Kazuma Kondo
Graduate School of Engineering, University of Fukui, Fukui, Japan
1 shared publication - Koki Takenaka
Graduate School of Engineering, University of Fukui, Fukui, Japan
1 shared publication - Makoto Koshino
National Institute of Technology, Ishikawa College, Tsubata 929-0392, Japan
1 shared publication Satoshi Kobayashi
Graduate School of Engineering, University of Fukui, Fukui, Japan
1 shared publication
Latest Publications
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11 (19) , 30978-30995
Expanding the Horizons of 1-D Tasks: Leveraging 2-D Convolutional Neural Networks Pretrained by ImageNet
Sensor-based human activity recognition (HAR) plays a pivotal role in the Internet of Things (IoT). Due to the low cost, low power consumption, and broad applicability of sensor devices, the integration of HAR with IoT has become increasingly imminent. Deep learning (DL) based HAR methods, with their capability for automatic feature extraction, are applied in our daily lives. However, obtaining a substantial amount of labeled data for HAR tasks remains a challenge. Although transfer learning (TL) offers a viable solution in the HAR domain, we lack large-source domain datasets akin to ImageNet. In response to this challenge, our study introduces an innovative cross-modal TL strategy. By compressing ImageNet parameters, which has allowed us to transfer the robust capabilities of 2D models into the 1D domain. This was achieved by employing various DL architectures, thereby validating the robustness of our method. Furthermore, by introducing the hyperparameter a and compressing the weight magnitude in TL, we further investigated our proposed method to enhance its universality. This is evidenced by the positive results obtained from HAR benchmark datasets. We also expanded our analysis to include results from different sensor types, affirming the adaptability and effectiveness of our TL strategy across diverse contexts. These experiments have not only reinforced our initial findings but have also widened the understanding of how our approach can be applied to a vast array of IoT-based scenarios and sensor data, opening up new avenues for research and application in the field.
- 17 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 11 (19) , 30978-30995
Expanding the Horizons of 1-D Tasks: Leveraging 2-D Convolutional Neural Networks Pretrained by ImageNet
Sensor-based human activity recognition (HAR) plays a pivotal role in the Internet of Things (IoT). Due to the low cost, low power consumption, and broad applicability of sensor devices, the integration of HAR with IoT has become increasingly imminent. Deep learning (DL) based HAR methods, with their capability for automatic feature extraction, are applied in our daily lives. However, obtaining a substantial amount of labeled data for HAR tasks remains a challenge. Although transfer learning (TL) offers a viable solution in the HAR domain, we lack large-source domain datasets akin to ImageNet. In response to this challenge, our study introduces an innovative cross-modal TL strategy. By compressing ImageNet parameters, which has allowed us to transfer the robust capabilities of 2D models into the 1D domain. This was achieved by employing various DL architectures, thereby validating the robustness of our method. Furthermore, by introducing the hyperparameter a and compressing the weight magnitude in TL, we further investigated our proposed method to enhance its universality. This is evidenced by the positive results obtained from HAR benchmark datasets. We also expanded our analysis to include results from different sensor types, affirming the adaptability and effectiveness of our TL strategy across diverse contexts. These experiments have not only reinforced our initial findings but have also widened the understanding of how our approach can be applied to a vast array of IoT-based scenarios and sensor data, opening up new avenues for research and application in the field.
- Open Access
- 17 February 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 20547-20558
A Comparative Study: Toward an Effective Convolutional Neural Network Architecture for Sensor-Based Human Activity Recognition
The feature extraction of human activity recognition (HAR) based on sensor data has been studied as a hand-crafted method. The significant feature extraction ability is a key factor in improving the accuracy of HAR. Recently, deep learning methods have been employed for feature extraction. In this paper, we review previous studies on deep learning methods in HAR and discuss suitable models for feature extraction. First, we applied various convolutional neural networks to clarify the effective architecture for HAR. Afterward, we developed advanced models by embedding submodules, such as self-attention and recurrent neural networks, often adopted in recent studies. Comparative experiments on HASC, UCI, and WISDM public datasets showed that Inception-V3, which used cross-channel multi-size convolution transformation, outperformed other backbones. Through comparative experiments after embedding submodules, submodules do not always have a positive effect on accuracy. Compared with other submodules, SENet has a positive effect. We conclude that it is essential to select an appropriate backbone model before applying the submodules, and submodules are unnecessary in some cases.
- 17 February 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 10, 20547-20558
Open AccessA Comparative Study: Toward an Effective Convolutional Neural Network Architecture for Sensor-Based Human Activity Recognition
The feature extraction of human activity recognition (HAR) based on sensor data has been studied as a hand-crafted method. The significant feature extraction ability is a key factor in improving the accuracy of HAR. Recently, deep learning methods have been employed for feature extraction. In this paper, we review previous studies on deep learning methods in HAR and discuss suitable models for feature extraction. First, we applied various convolutional neural networks to clarify the effective architecture for HAR. Afterward, we developed advanced models by embedding submodules, such as self-attention and recurrent neural networks, often adopted in recent studies. Comparative experiments on HASC, UCI, and WISDM public datasets showed that Inception-V3, which used cross-channel multi-size convolution transformation, outperformed other backbones. Through comparative experiments after embedding submodules, submodules do not always have a positive effect on accuracy. Compared with other submodules, SENet has a positive effect. We conclude that it is essential to select an appropriate backbone model before applying the submodules, and submodules are unnecessary in some cases.