ISSN 
Current Articles
2024, Volume 53, Issue 5
2024,
53(5):
641-654.
doi: 10.12178/1001-0548.2024196
Abstract:
With the rapid development of artificial intelligence technology, deep learning has shown tremendous potential in the field of spectral regulation of micro-nano structures. By constructing complex neural network models, deep learning can learn the spectral response characteristics of micro-nano structures from experimental or simulation data without the need for explicit physical analytical models, thereby achieving efficient design optimization. This provides a new approach and methodology for the design of micro-nano structures. This paper reviews the recent research progress of deep learning in micro-nano structure design, focusing on its applications in structural color, thermal radiation control, and narrowband spectral sensing, and also discusses future opportunities and challenges in this field.
With the rapid development of artificial intelligence technology, deep learning has shown tremendous potential in the field of spectral regulation of micro-nano structures. By constructing complex neural network models, deep learning can learn the spectral response characteristics of micro-nano structures from experimental or simulation data without the need for explicit physical analytical models, thereby achieving efficient design optimization. This provides a new approach and methodology for the design of micro-nano structures. This paper reviews the recent research progress of deep learning in micro-nano structure design, focusing on its applications in structural color, thermal radiation control, and narrowband spectral sensing, and also discusses future opportunities and challenges in this field.
2024,
53(5):
655-671.
doi: 10.12178/1001-0548.2024177
Abstract:
The technology of short-wave infrared photodetection has achieved important applications in military defense, industrial manufacturing, medical diagnosis and other fields, but the high price limits its popularization and application scenarios. Development of short-wave infrared photodetection based on new photosensitive materials is considered to be the key to the large-scale application of this technology. Organic photodetectors have significant advantages in terms of low cost and flexibility, and are developing rapidly in the field of short-wave infrared photodetection. It is expected to be a complementary technology to InGaAs and quantum dots based ones, with the ability to achieve large-scale applications on the low-cost demand side in the field of Internet of Things and artificial intelligence economy. In this paper, we first introduce the basic working principles and performance parameters of organic photodetectors, followed by the reviewing of short-wave infrared photodetection polymers and small molecules, and finally the applications of organic short-wave infrared photodetectors are summarized.
The technology of short-wave infrared photodetection has achieved important applications in military defense, industrial manufacturing, medical diagnosis and other fields, but the high price limits its popularization and application scenarios. Development of short-wave infrared photodetection based on new photosensitive materials is considered to be the key to the large-scale application of this technology. Organic photodetectors have significant advantages in terms of low cost and flexibility, and are developing rapidly in the field of short-wave infrared photodetection. It is expected to be a complementary technology to InGaAs and quantum dots based ones, with the ability to achieve large-scale applications on the low-cost demand side in the field of Internet of Things and artificial intelligence economy. In this paper, we first introduce the basic working principles and performance parameters of organic photodetectors, followed by the reviewing of short-wave infrared photodetection polymers and small molecules, and finally the applications of organic short-wave infrared photodetectors are summarized.
2024,
53(5):
672-684.
doi: 10.12178/1001-0548.2024188
Abstract:
Optical wavefront is the geometric representation of the phase surface of light waves. Optical wavefront sensing technology analyzes the properties of objects by detecting phase changes in the light wave propagation path. This technology is widely used in atmospheric turbulence detection, optical element defect analysis, and biological sample research, playing a crucial role in fields such as astronomy, adaptive optics, microscopic imaging, laser systems, and biomedicine. However, common detectors are only sensitive to light intensity. To detect optical wavefronts, a series of complex optical components are typically required at the detection front end, leading to large system sizes, high costs, and structural complexity. In recent years, with continuous advancements in micro-nano optics and artificial intelligence, a series of integrated, miniaturized, and high-performance optical wavefront sensing technologies based on new principles, devices, and algorithms have emerged. This paper systematically reviews the recent research progress in optical wavefront sensing techniques, including two main types of interferometric and non-interferometric as well as typical methods: shear interferometry type, grating interferometry type, near-field interferometry type, algorithmic reconstruction type, and dimension-associated type. Finally, the current challenges in the field are summarized and the future development directions prospected.
Optical wavefront is the geometric representation of the phase surface of light waves. Optical wavefront sensing technology analyzes the properties of objects by detecting phase changes in the light wave propagation path. This technology is widely used in atmospheric turbulence detection, optical element defect analysis, and biological sample research, playing a crucial role in fields such as astronomy, adaptive optics, microscopic imaging, laser systems, and biomedicine. However, common detectors are only sensitive to light intensity. To detect optical wavefronts, a series of complex optical components are typically required at the detection front end, leading to large system sizes, high costs, and structural complexity. In recent years, with continuous advancements in micro-nano optics and artificial intelligence, a series of integrated, miniaturized, and high-performance optical wavefront sensing technologies based on new principles, devices, and algorithms have emerged. This paper systematically reviews the recent research progress in optical wavefront sensing techniques, including two main types of interferometric and non-interferometric as well as typical methods: shear interferometry type, grating interferometry type, near-field interferometry type, algorithmic reconstruction type, and dimension-associated type. Finally, the current challenges in the field are summarized and the future development directions prospected.
2024,
53(5):
685-697.
doi: 10.12178/1001-0548.2024225
Abstract:
The superior material properties with wide bandgap, large critical electric field, and high saturated electron velocity, in combination with the high density and high mobility two-dimensional electron gas induced at the AlGaN/GaN heterojunction by polarization discontinuity, and thus the related high electron mobility transistors, make GaN devices become new high performance electronic devices for next-generation power and RF applications. The demand for GaN-based power devices with excellent performance in emerging technology such as electric vehicles and AI is rapidly increasing. GaN single-chip power integration technology is the key approach to reduce the influence of parasitic inductance, improve the switching speed of IC, cut down the power consumption and realize the miniaturization for the whole system. Based on GaN single-chip power integration technology, this review paper presents a comprehensive and global overview for the research progress of the reported double-heterojunction based epitaxial structure with p-/n-channels, monolithic heterogeneous integration, All-GaN integrated circuits, and the core technology of p-channel devices.
The superior material properties with wide bandgap, large critical electric field, and high saturated electron velocity, in combination with the high density and high mobility two-dimensional electron gas induced at the AlGaN/GaN heterojunction by polarization discontinuity, and thus the related high electron mobility transistors, make GaN devices become new high performance electronic devices for next-generation power and RF applications. The demand for GaN-based power devices with excellent performance in emerging technology such as electric vehicles and AI is rapidly increasing. GaN single-chip power integration technology is the key approach to reduce the influence of parasitic inductance, improve the switching speed of IC, cut down the power consumption and realize the miniaturization for the whole system. Based on GaN single-chip power integration technology, this review paper presents a comprehensive and global overview for the research progress of the reported double-heterojunction based epitaxial structure with p-/n-channels, monolithic heterogeneous integration, All-GaN integrated circuits, and the core technology of p-channel devices.
2024,
53(5):
698-705.
doi: 10.12178/1001-0548.2024212
Abstract:
To enhance the comprehensive prevention, control, and emergency response capabilities for forest and grassland fires, aiming to reduce the occurrence of such fires and achieve rapid dynamic monitoring of fire situations, the keypoint is lying on developing the information-based intelligent theories and methods for fire monitoring and early warning by integrating remote sensing, geospatial information, computer technology, fire ecology, and other disciplines. Based on the methods including remote sensing quantitative inversion of fuel variables and spatiotemporal big data mining, the technology system of remote sensing monitoring and early warning for forest and grassland fires for the pre-fire/during fire/post-fire period is established. It mainly includes the fire risk forecasting and early warning technology incorporating multisource spatiotemporal big data on fuel, weather and topography variables, the rapid and high-precision fire detection technology based on multisource satellite data, and the technology of accurate burn severity estimations. In response to the practical applications in southwestern China, the customized system for forest and grassland fire monitoring and early warning is developed, involving the key methodologies and technologies of fire monitoring and early warning. This system could achieve functions of fire risk early warning, near-real-time fire monitoring and fast assessment of fire loss. It also provides convenient and efficient dynamic information services for forest and grassland fire monitoring and early warning, and shows the significant practical application effectiveness, indicating the potential for its widespread adoption.
To enhance the comprehensive prevention, control, and emergency response capabilities for forest and grassland fires, aiming to reduce the occurrence of such fires and achieve rapid dynamic monitoring of fire situations, the keypoint is lying on developing the information-based intelligent theories and methods for fire monitoring and early warning by integrating remote sensing, geospatial information, computer technology, fire ecology, and other disciplines. Based on the methods including remote sensing quantitative inversion of fuel variables and spatiotemporal big data mining, the technology system of remote sensing monitoring and early warning for forest and grassland fires for the pre-fire/during fire/post-fire period is established. It mainly includes the fire risk forecasting and early warning technology incorporating multisource spatiotemporal big data on fuel, weather and topography variables, the rapid and high-precision fire detection technology based on multisource satellite data, and the technology of accurate burn severity estimations. In response to the practical applications in southwestern China, the customized system for forest and grassland fire monitoring and early warning is developed, involving the key methodologies and technologies of fire monitoring and early warning. This system could achieve functions of fire risk early warning, near-real-time fire monitoring and fast assessment of fire loss. It also provides convenient and efficient dynamic information services for forest and grassland fire monitoring and early warning, and shows the significant practical application effectiveness, indicating the potential for its widespread adoption.
2024,
53(5):
706-719.
doi: 10.12178/1001-0548.2024165
Abstract:
With advantages such as small footprint, low power consumption, reconfigurability and low cost, chip-integrated coherent light sources have become a highly promising scheme, applied in optical computing, autonomous driving, high-speed optical interconnection, 5G/6G communications, etc. Combining on-chip integration and cavity enhancement effects, microcavity-based integrated coherent light sources significantly improve the intracavity light power as well as reducing the pump threshold. This paper introduces in detail the principle and research progress of four kinds of microcavity integrated coherent light sources: Kerr frequency combs, microcavity stimulated Raman lasers, quadratic combs and microcavity-enhanced electro-optical combs. The four key parameters of spectrum width, noise, efficiency and power are analyzed one by one, and the research progresses of microcavity-based integrated coherent light sources in communication applications are introduced. At last their future development trends are prospected.
With advantages such as small footprint, low power consumption, reconfigurability and low cost, chip-integrated coherent light sources have become a highly promising scheme, applied in optical computing, autonomous driving, high-speed optical interconnection, 5G/6G communications, etc. Combining on-chip integration and cavity enhancement effects, microcavity-based integrated coherent light sources significantly improve the intracavity light power as well as reducing the pump threshold. This paper introduces in detail the principle and research progress of four kinds of microcavity integrated coherent light sources: Kerr frequency combs, microcavity stimulated Raman lasers, quadratic combs and microcavity-enhanced electro-optical combs. The four key parameters of spectrum width, noise, efficiency and power are analyzed one by one, and the research progresses of microcavity-based integrated coherent light sources in communication applications are introduced. At last their future development trends are prospected.
2024,
53(5):
720-731.
doi: 10.12178/1001-0548.2024176
Abstract:
Emotion recognition is an important research direction in the fields of artificial intelligence and human-computer interaction. It has significant implications for enhancing user experience and the intelligence of applications. Emotion recognition based on multimodal physiological data has become a research hotspot in recent years due to the objectivity and diversity of its data sources, which enable more accurate capture of an individual's emotional state. Firstly, the basic concepts of affective computing and emotion representation models are introduced. Secondly, emotion recognition methods based on physiological data are summarized. Then, the focus shifts to the process of emotion recognition based on multimodal physiological data, including physiological data preprocessing, traditional machine learning methods, and deep learning methods. Finally, the main challenges faced by emotion recognition based on multimodal physiological data are analyzed, and future prospects are discussed.
Emotion recognition is an important research direction in the fields of artificial intelligence and human-computer interaction. It has significant implications for enhancing user experience and the intelligence of applications. Emotion recognition based on multimodal physiological data has become a research hotspot in recent years due to the objectivity and diversity of its data sources, which enable more accurate capture of an individual's emotional state. Firstly, the basic concepts of affective computing and emotion representation models are introduced. Secondly, emotion recognition methods based on physiological data are summarized. Then, the focus shifts to the process of emotion recognition based on multimodal physiological data, including physiological data preprocessing, traditional machine learning methods, and deep learning methods. Finally, the main challenges faced by emotion recognition based on multimodal physiological data are analyzed, and future prospects are discussed.
2024,
53(5):
732-748.
doi: 10.12178/1001-0548.2023260
Abstract:
In recent years, with the rapid development of deep learning technology, the research of person search algorithms has attracted a lot of scholars' interest. Person search is to find specific target person in images based on person detection and person re-identification tasks. In this paper, we review the recent research progress on person search task in detail. The existing methods are analyzed and summarized in terms of model network structures and loss functions. According to the two different technical routes of convolutional neural network and Transformer, the main research work of their respective representative methods is focused on. According to the traditional loss function, OIM loss function, and mixed loss function, the training loss functions used in person search are summarized. In addition, the public data sets commonly used in the field of person search are summarized, and the performances of the main algorithms on the corresponding data sets are compared and analyzed. Finally, we summarize the future research directions of person search task.
In recent years, with the rapid development of deep learning technology, the research of person search algorithms has attracted a lot of scholars' interest. Person search is to find specific target person in images based on person detection and person re-identification tasks. In this paper, we review the recent research progress on person search task in detail. The existing methods are analyzed and summarized in terms of model network structures and loss functions. According to the two different technical routes of convolutional neural network and Transformer, the main research work of their respective representative methods is focused on. According to the traditional loss function, OIM loss function, and mixed loss function, the training loss functions used in person search are summarized. In addition, the public data sets commonly used in the field of person search are summarized, and the performances of the main algorithms on the corresponding data sets are compared and analyzed. Finally, we summarize the future research directions of person search task.
2024,
53(5):
749-761.
doi: 10.12178/1001-0548.2024171
Abstract:
Accurate state estimation and prediction of lithium-ion battery are crucial for ensuring operational performance and safety. Data-driven state estimation algorithms are prone to the distribution shift between training data and testing data, limiting their generalization capabilities. Transfer-learning-based cross-domain state estimation algorithms are proposed to address these issues. This paper discusses around three common application scenarios: state of charge estimation, state of health estimation, and remaining useful life estimation. While comparing the differences between methods across various scenarios, the review also reveals their commonalities. From a technical perspective, this paper categorizes commonly used transfer methods into three types: finetuning-based transfer, metric-based transfer, and adversarial training-based transfer. Based on these technical approaches, this paper provides a comprehensive and clear summary of recent cross-domain lithium-ion battery state estimation methods.
Accurate state estimation and prediction of lithium-ion battery are crucial for ensuring operational performance and safety. Data-driven state estimation algorithms are prone to the distribution shift between training data and testing data, limiting their generalization capabilities. Transfer-learning-based cross-domain state estimation algorithms are proposed to address these issues. This paper discusses around three common application scenarios: state of charge estimation, state of health estimation, and remaining useful life estimation. While comparing the differences between methods across various scenarios, the review also reveals their commonalities. From a technical perspective, this paper categorizes commonly used transfer methods into three types: finetuning-based transfer, metric-based transfer, and adversarial training-based transfer. Based on these technical approaches, this paper provides a comprehensive and clear summary of recent cross-domain lithium-ion battery state estimation methods.
2024,
53(5):
762-770.
doi: 10.12178/1001-0548.2024173
Abstract:
As globalization continues to develop, cross-lingual summarization has become an important topic in natural language processing. In low-resource scenarios, existing methods face challenges such as limited representation transfer and insufficient data utilization. To address these issues, this paper proposes a novel method based on joint training and self-training. Specifically, two models are used to handle the translation and cross-lingual summarization tasks, respectively, which unify the language vector space of the output and avoid the issue of limited representation transfer. Additionally, joint training is performed by aligning the output features and probabilities of parallel training pairs, thereby enhancing semantic sharing between the models. Furthermore, based on joint training, a self-training technique is introduced to generate synthetic data from additional monolingual summary data, effectively mitigating the data scarcity issue of low-resource scenarios. Experimental results demonstrate that this method outperforms existing approaches in multiple low-resource scenarios, achieving significant improvements in ROUGE scores.
As globalization continues to develop, cross-lingual summarization has become an important topic in natural language processing. In low-resource scenarios, existing methods face challenges such as limited representation transfer and insufficient data utilization. To address these issues, this paper proposes a novel method based on joint training and self-training. Specifically, two models are used to handle the translation and cross-lingual summarization tasks, respectively, which unify the language vector space of the output and avoid the issue of limited representation transfer. Additionally, joint training is performed by aligning the output features and probabilities of parallel training pairs, thereby enhancing semantic sharing between the models. Furthermore, based on joint training, a self-training technique is introduced to generate synthetic data from additional monolingual summary data, effectively mitigating the data scarcity issue of low-resource scenarios. Experimental results demonstrate that this method outperforms existing approaches in multiple low-resource scenarios, achieving significant improvements in ROUGE scores.
2024,
53(5):
771-784.
doi: 10.12178/1001-0548.2024215
Abstract:
Emotion recognition, endowing computers with the ability to perceive emotions, is a focal point of interest in various fields, including computer science, psychology, sociology, biomedical engineering, and more. EEG network analysis methods are widely used in the neuroimaging field for neurocognitive analysis. These methods capture interactions between/among different brain regions to construct brain networks, thereby describing the information flow and functional collaboration across various brain areas. Given that emotional functions inherently involve the cooperation of multiple brain regions, EEG network analysis methods excel in capturing inter-regional information interactions, making them highly effective in emotion recognition. This paper provides a comprehensive introduction to the research background, principles, methods, and current status of EEG network-based emotion recognition. Additionally, the existing challenges and future development trends in this research area are discussed.
Emotion recognition, endowing computers with the ability to perceive emotions, is a focal point of interest in various fields, including computer science, psychology, sociology, biomedical engineering, and more. EEG network analysis methods are widely used in the neuroimaging field for neurocognitive analysis. These methods capture interactions between/among different brain regions to construct brain networks, thereby describing the information flow and functional collaboration across various brain areas. Given that emotional functions inherently involve the cooperation of multiple brain regions, EEG network analysis methods excel in capturing inter-regional information interactions, making them highly effective in emotion recognition. This paper provides a comprehensive introduction to the research background, principles, methods, and current status of EEG network-based emotion recognition. Additionally, the existing challenges and future development trends in this research area are discussed.
2024,
53(5):
785-791.
doi: 10.12178/1001-0548.2023121
Abstract:
In order to alleviate traffic congestion on freeways, national highways and provincial highways are used to achieve traffic diversion, and a targeted route guidance strategy is implemented. A hybrid routing model for regional multi-layer transportation network is established, which considers the congested driver sources. Firstly, a regional multi-layer transportation network composed of Hunan freeway, national highway and provincial highway is constructed, and the link weight is calibrated. Next, the congested driver sources are located, a hybrid routing model is established by considering the congested driver sources, and the targeted routing guidance is implemented for vehicles from the congested driver sources, so as to reduce the impact on travelers as much as possible while ensuring the mitigation effect of traffic congestion. Results show that the proposed route guidance model can effectively achieve traffic diversion and reduce the average travel time of vehicles in the multi-layer transportation network. In practical applications, the route guidance information can be only sent to a few drivers to reduce the difficulty of model implementation. The acceptance rate of drivers to the suggested routes can be analyzed to develop appropriate plans for disseminating route guidance information.
In order to alleviate traffic congestion on freeways, national highways and provincial highways are used to achieve traffic diversion, and a targeted route guidance strategy is implemented. A hybrid routing model for regional multi-layer transportation network is established, which considers the congested driver sources. Firstly, a regional multi-layer transportation network composed of Hunan freeway, national highway and provincial highway is constructed, and the link weight is calibrated. Next, the congested driver sources are located, a hybrid routing model is established by considering the congested driver sources, and the targeted routing guidance is implemented for vehicles from the congested driver sources, so as to reduce the impact on travelers as much as possible while ensuring the mitigation effect of traffic congestion. Results show that the proposed route guidance model can effectively achieve traffic diversion and reduce the average travel time of vehicles in the multi-layer transportation network. In practical applications, the route guidance information can be only sent to a few drivers to reduce the difficulty of model implementation. The acceptance rate of drivers to the suggested routes can be analyzed to develop appropriate plans for disseminating route guidance information.
2024,
53(5):
792-800.
doi: 10.12178/1001-0548.2024205
Abstract:
Link prediction is one of the most productive branches in network science, aiming to estimate the likelihoods of unobserved links based on known network topology. This paper critically examines four fundamental issues in link prediction, say network selection, link sampling, model training and algorithm evaluation. It reviews the current research progresses and highlights some significant yet unresolved issues that urgently require scientific answers.
Link prediction is one of the most productive branches in network science, aiming to estimate the likelihoods of unobserved links based on known network topology. This paper critically examines four fundamental issues in link prediction, say network selection, link sampling, model training and algorithm evaluation. It reviews the current research progresses and highlights some significant yet unresolved issues that urgently require scientific answers.
