Kai Song
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Kai Song has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Dong Wang
Taiyuan University of Technology, College of Physics and Optoelectronics, Yingze, China
1 shared publication - Fanjin Zeng
College of Physics and Optoelectronics Engineering, Taiyuan University of Technology 1 , Taiyuan 030024
1 shared publication - Futong Zhang
College of Physics and Optoelectronics Engineering, Taiyuan University of Technology 1 , Taiyuan 030024
1 shared publication - Hongda Ge
College of Physics and Optoelectronics Engineering, Taiyuan University of Technology 1 , Taiyuan 030024
1 shared publication - Hongrui Liu
College of Physics and Optoelectronics Engineering, Taiyuan University of Technology 1 , Taiyuan 030024
1 shared publication - Juanjuan Zhao
Taiyuan University of Technology
1 shared publication
Latest Publications
- Open Access
- 1 June 2025
- Published by AIP Publishing in APL Photonics
- Vol. 10 (6)
Photon-level single-pixel wavefront imaging through turbid underwater environment
Single-pixel wavefront imaging provides the possibility to accurately measure and process complex wavefronts in extremely low-light scenes. However, the noises introduced in the turbid environment greatly affect the quality of wavefront imaging. Here, we propose a highly robust single-pixel wavefront detection technology with phase-denoising neural network. Aiming at different levels of noise in various turbid underwater environments, the proposed technology demonstrates adaptive denoising capabilities and can actively adjust to variations in underwater environments. Based on the advantages of single-photon detection technology, the wavefront images with arbitrary orbital angular momentum modes can be reconstructed with only 0.022 photon per pattern per pixel. Meanwhile, the results demonstrate that the proposed technology can effectively reduce noises in underwater environments with turbidity levels ranging from 0.1 to 100 NTU. The presented technology promotes the practical application of single-pixel wavefront imaging systems in the underwater visible light communication.
- 1 June 2025
- Published by AIP Publishing in APL Photonics
- Vol. 10 (6)
Open AccessPhoton-level single-pixel wavefront imaging through turbid underwater environment
Single-pixel wavefront imaging provides the possibility to accurately measure and process complex wavefronts in extremely low-light scenes. However, the noises introduced in the turbid environment greatly affect the quality of wavefront imaging. Here, we propose a highly robust single-pixel wavefront detection technology with phase-denoising neural network. Aiming at different levels of noise in various turbid underwater environments, the proposed technology demonstrates adaptive denoising capabilities and can actively adjust to variations in underwater environments. Based on the advantages of single-photon detection technology, the wavefront images with arbitrary orbital angular momentum modes can be reconstructed with only 0.022 photon per pattern per pixel. Meanwhile, the results demonstrate that the proposed technology can effectively reduce noises in underwater environments with turbidity levels ranging from 0.1 to 100 NTU. The presented technology promotes the practical application of single-pixel wavefront imaging systems in the underwater visible light communication.
- Open Access
- 26 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 23-33
Two-domain adaptive attention network for TCM auxiliary diagnosis with hand images
Cerebral Stroke (CS) has now become the second leading cause of death in the world, with extremely high fatality and disability rates, often bringing irreversible serious consequences to patients. Due to the lack of effective treatment, prevention is currently considered the best measure. The early diagnosis of CS relies on professional medical techniques, such as electroencephalography, cerebral angiography, etc., which usually require expensive prices and professional techniques, and even bring great pain to patients. Considering the many inconveniences and risks of existing traditional diagnostic methods, and inspired by the idea of Traditional Chinese Medicine(TCM) inspection, this paper proposes a deep learning method for CS recognition using hand images based on computer image processing technology, named dual-domain adaptive attention network. First, a general backbone network is constructed to learn various information contained in palm images. Then, an area adaptive module and a channel adaptive module are proposed to learn the contribution of different palm regions and different angle features to the results. Next, the contribution degree is used to guide the hybrid-domain attention module to further learn the correlation between key features to ensure that the model learns as many and useful features as possible. The experimental results showed that the recognition accuracy of our model for CS can reach 86.81%, the average sensitivity is 87.38%, the average specificity is 85.98%, and the average area under the curve (AUC) is 93.78%, which is significantly improved compared with traditional classification methods. this paper also verifies the feasibility of automatic disease recognition using palm images, providing the intelligence of traditional Chinese medicine a new idea.
- 26 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 23-33
Open AccessTwo-domain adaptive attention network for TCM auxiliary diagnosis with hand images
Cerebral Stroke (CS) has now become the second leading cause of death in the world, with extremely high fatality and disability rates, often bringing irreversible serious consequences to patients. Due to the lack of effective treatment, prevention is currently considered the best measure. The early diagnosis of CS relies on professional medical techniques, such as electroencephalography, cerebral angiography, etc., which usually require expensive prices and professional techniques, and even bring great pain to patients. Considering the many inconveniences and risks of existing traditional diagnostic methods, and inspired by the idea of Traditional Chinese Medicine(TCM) inspection, this paper proposes a deep learning method for CS recognition using hand images based on computer image processing technology, named dual-domain adaptive attention network. First, a general backbone network is constructed to learn various information contained in palm images. Then, an area adaptive module and a channel adaptive module are proposed to learn the contribution of different palm regions and different angle features to the results. Next, the contribution degree is used to guide the hybrid-domain attention module to further learn the correlation between key features to ensure that the model learns as many and useful features as possible. The experimental results showed that the recognition accuracy of our model for CS can reach 86.81%, the average sensitivity is 87.38%, the average specificity is 85.98%, and the average area under the curve (AUC) is 93.78%, which is significantly improved compared with traditional classification methods. this paper also verifies the feasibility of automatic disease recognition using palm images, providing the intelligence of traditional Chinese medicine a new idea.