Article
Optics
Shoupei Liu, Xiangfeng Meng, Yongkai Yin, Huazheng Wu, Wenjie Jiang
Summary: UNNCGI is a computational ghost imaging method that employs an untrained neural network to generate high-quality images even at low sampling ratios, improving imaging efficiency.
OPTICS AND LASERS IN ENGINEERING
(2021)
Article
Engineering, Electrical & Electronic
Jorge Bacca, Tatiana Gelvez-Barrera, Henry Arguello
Summary: Covering a wide range of CI problems, the end-to-end deep learning-based optimization of CAs aims to easily change the loss function of the deep approach and includes regularizers to fulfill the widely used sensing requirements of the CI applications. The binary CA solution is encouraged, and the performance of the CI task is maximized in applications such as restoration, classification, and semantic segmentation.
IEEE TRANSACTIONS ON COMPUTATIONAL IMAGING
(2021)
Article
Optics
Wenhan Ren, Xiaoyu Nie, Tao Peng, Marlan O. Scully
Summary: Artificial intelligence is widely used in computational imaging to improve the quality and signal-to-noise ratio of images affected by low sampling ratio or noisy environments. This study proposes a new computational imaging scheme based on the sequence transduction mechanism with the transformer network, which can retrieve high-quality images and is robust to noise interference.
Article
Physics, Applied
Xiangyu Zhang, Fei Wang, Guohai Situ
Summary: The solution of an inverse problem in computational imaging often relies on accurate knowledge of the physical model and object, but model uncertainty in practical applications can degrade the quality of reconstructed images. In this paper, we propose a novel untrained learning approach to address computational imaging with model uncertainty, demonstrated through phase retrieval, an important task in biomedical imaging and industrial inspection.
JOURNAL OF PHYSICS D-APPLIED PHYSICS
(2022)
Article
Engineering, Electrical & Electronic
Jialong Xu, Bo Ai, Ning Wang, Wei Chen
Summary: This research proposes a deep joint source-channel coding (DJSCC) based framework for the CSI feedback task in MIMO technology. The proposed framework combines non-linear transform networks to compress the CSI and an SNR adaption mechanism to deal with wireless channel variations, resulting in improved performance in limited bandwidth and low SNR environments.
IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS
(2023)
Article
Physics, Multidisciplinary
Yi-Yi Huang, Chen Ou-Yang, Ke Fang, Yu-Feng Dong, Jie Zhang, Li-Ming Chen, Ling-An Wu
Summary: An overlapping sampling scheme is proposed to accelerate computational ghost imaging for moving targets. The method reorders Hadamard modulation matrices and uses deep learning to improve imaging speed and quality by training a neural network with acquired signals and real images from bucket detectors. Detailed comparisons demonstrate a significant improvement in imaging speed and quality using this new approach.
Article
Optics
Leihong Zhang, Yunjie Zhai, Runchu Xu, Kaimin Wang, Dawei Zhang
Summary: This paper investigates the impact of atmospheric turbulence on the image acquisition process and uses computational ghost imaging to simulate the process. The study finds that good reconstruction results can be obtained using an end-to-end neural network at low sampling rates and extreme conditions.
Article
Optics
Zhan Yu, Xinjia Li, Xing Bai, Yujie Wang, Xingyu Chen, Yang Liu, Mingze Sun, Xin Zhou
Summary: This article presents an imaging method through a dynamic scattering medium based on computational ghost imaging and a convolutional neural network. The CNN is used to improve the quality of ghost imaging, and its training set is obtained from numerical simulation rather than actual experiments, reducing the workload significantly. A concise mathematical model is provided to reflect the absorption and scattering effects of the dynamic medium. By adding Gaussian white noise to the detected light intensity sequence, the undulation caused by the dynamic scatterer is simulated, and the network is trained under these conditions. Compared to the dataset without noise, our proposed method demonstrates better performance in imaging through a dynamic scattering medium, for both simple binary objects and complex grayscale ones. The effectiveness of this method has been verified in experiments with scattering medium rotated at different speeds.
LASER PHYSICS LETTERS
(2023)
Article
Multidisciplinary Sciences
Chane Moodley, Bereneice Sephton, Valeria Rodriguez-Fajardo, Andrew Forbes
Summary: A two-step deep learning approach has been proposed for optimizing image reconstruction and determining an accurate early stopping point in quantum ghost imaging, resulting in a significant decrease in image acquisition time. This method has also achieved a notable reduction in experimental running time.
SCIENTIFIC REPORTS
(2021)
Article
Engineering, Civil
Luca Anzalone, Paola Barra, Silvio Barra, Aniello Castiglione, Michele Nappi
Summary: This work combines Curriculum Learning with Deep Reinforcement Learning to learn a competitive driving policy without prior domain knowledge in the CARLA autonomous driving simulator. The approach divides the reinforcement learning phase into multiple stages of increasing difficulty, guiding the agent towards an increasingly better driving policy. The agent architecture includes various neural networks and novel value decomposition scheme and gradient size normalization function. Quantitative and qualitative results of the learned driving policy are presented.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2022)
Article
Construction & Building Technology
Chen Wang, Ling-han Song, Jian-sheng Fan
Summary: This paper presents DeepSNA, a general computational framework in civil engineering that can predict the mechanical responses of different structures based on deep learning. The framework considers both intrinsic structural information and external excitations, and achieves high accuracy and computational efficiency through careful design of data interface schema, deep learning models, and data augmentation algorithms.
AUTOMATION IN CONSTRUCTION
(2022)
Article
Computer Science, Interdisciplinary Applications
Hans Pinckaers, Wouter Bulten, Jeroen van der Laak, Geert Litjens
Summary: Prostate cancer is the most prevalent cancer among men in Western countries, and pathologists' evaluation is the gold standard for diagnosis. State-of-the-art convolutional neural networks are often patch-based and require detailed pixel-level annotations for effective training.
IEEE TRANSACTIONS ON MEDICAL IMAGING
(2021)
Article
Optics
Xuanpengfan Zou, Xianwei Huang, Cong Liu, Wei Tan, Yanfeng Bai, Xiquan Fu
Summary: Recognizing the target of interest in an imaging area is crucial. Ghost imaging (GI) shows potential in target recognition by presenting an anti-disturbance property. This paper proposes a characteristic imaging model for target recognition based on deep learning in computational GI. The results demonstrate the potential applications of this imaging scheme in recognizing the target of interest.
OPTICS AND LASER TECHNOLOGY
(2023)
Article
Physics, Multidisciplinary
Hao Zhang, Yunjie Xia, Deyang Duan
Summary: In this study, a novel approach using a deep convolutional generative adversarial network for compressed sensing algorithm is proposed to enhance the imaging performance of computational ghost imaging. The results show significant improvement in image quality and effective noise elimination.
Article
Agriculture, Dairy & Animal Science
Chengyun Zhang, Yonghuan Chen, Zezhou Hao, Xinghui Gao
Summary: This study proposes an efficient time-domain single-channel bird sound separation network that achieves good separation performance and high separation efficiency. By utilizing massive amounts of bird sound data and incorporating a dual-path network and simplified transformer structure, the network significantly reduces computational resources while maintaining good separation results. The proposed network has the potential to contribute to distinguishing individual birds, studying bird interactions, and enabling automatic identification of bird species on various mobile and edge computing devices.
Article
Optics
Joseph Rosen, Hilton B. de Aguiar, Vijayakumar Anand, YoonSeok Baek, Sylvain Gigan, Ryoichi Horisaki, Herve Hugonnet, Saulius Juodkazis, KyeoReh Lee, Haowen Liang, Yikun Liu, Stephan Ludwig, Wolfgang Osten, YongKeun Park, Giancarlo Pedrini, Tushar Sarkar, Johannes Schindler, Alok Kumar Singh, Rakesh Kumar Singh, Guohai Situ, Mitsuo Takeda, Xiangsheng Xie, Wanqin Yang, Jianying Zhou
Summary: In recent years, the rapid development of chaos-inspired imaging technologies, consisting of non-invasive and invasive directions, has led to faster and smarter imaging capabilities. Non-invasive imaging through scattering layers has achieved significant progress, while invasive imaging exploits chaos to achieve special imaging characteristics and increase dimensionalities beyond the limits of conventional imagers. This roadmap presents the current and future challenges in both invasive and non-invasive imaging technologies.
APPLIED PHYSICS B-LASERS AND OPTICS
(2022)
Article
Physics, Applied
Xiangyu Zhang, Fei Wang, Guohai Situ
Summary: The solution of an inverse problem in computational imaging often relies on accurate knowledge of the physical model and object, but model uncertainty in practical applications can degrade the quality of reconstructed images. In this paper, we propose a novel untrained learning approach to address computational imaging with model uncertainty, demonstrated through phase retrieval, an important task in biomedical imaging and industrial inspection.
JOURNAL OF PHYSICS D-APPLIED PHYSICS
(2022)
Article
Optics
Fei Wang, Chenglong Wang, Mingliang Chen, Wenlin Gong, Yu Zhang, Shensheng Han, Guohai Situ
Summary: Ghost imaging (GI) is a technique that facilitates image acquisition under low-light conditions by single-pixel measurements, and has great potential in various applications. However, the current GI techniques require a large number of samples to reconstruct high-resolution images, which limits their practical applications. In this study, a far-field super-resolution GI technique is proposed that incorporates the physical model for GI image formation into a deep neural network. The proposed technique outperforms other widespread GI techniques in terms of spatial resolution and sampling ratio, and provides a new framework for GI with practical applications.
LIGHT-SCIENCE & APPLICATIONS
(2022)
Article
Optics
Fei Wang, Chenglong Wang, Chenjin Deng, Shensheng Han, Guohai Situ
Summary: This study proposes a physics-enhanced deep learning approach for image reconstruction in single-pixel imaging. By combining a physics-informed layer and a model-driven fine-tuning process, the proposed method demonstrates generalizability and outperforms other widespread algorithms in terms of both robustness and fidelity.
PHOTONICS RESEARCH
(2022)
Article
Optics
Dong Liang, Xiaoting Peng, Yuyao Hu, Fu Zhao, Shanshan Zheng, Guohai Situ, Jun Liu
Summary: Light sheet fluorescence microscopy (LSFM) using light field imaging can improve axial resolution and provide high axial resolution and large field of view without losing spatial resolution.
OPTICS AND LASERS IN ENGINEERING
(2022)
Article
Optics
Dajiang Lu, Qi Xing, Meihua Liao, Guohai Situ, Xiang Peng, Wenqi He
Summary: In this study, we experimentally investigate image reconstruction through a scattering medium under white-light illumination. A modified iterative algorithm with a constraint on the optical transfer function (OTF) is used to solve the inverse problem of noninvasive scattering imaging. The proposed method shows a potential advantage in scattering imaging compared to the well-known speckle correlation technique (SCT).
Article
Optics
Yong-Liang Xiao, Sikun Li, Guohai Situ, Jianxin Zhong
Summary: This Letter presents a unitary neural network based on optical random phase DropConnect, which introduces a micro-phase to achieve training convergence and enhance statistical inference. The study reveals that the partial drilling of random micro-phase-shift outperforms the full-drilling counterpart in both training and inference.
Article
Nanoscience & Nanotechnology
Xiangyu Zhang, Chenjin Deng, Chenglong Wang, Fei Wang, Guohai Situ
Summary: Single-pixel imaging (SPI) converts a multi-dimensional image acquisition problem into a one-dimensional temporal-signal detection problem, and efficient SPI techniques are crucial for image reconstruction. Deep learning has shown superiority in SPI, but is task-specific and requires retraining for different problems. The proposed VGenNet algorithm incorporates a model-driven fine-tuning process into a generative model, allowing the use of a pretrained model for various inverse imaging problems. Experimental results demonstrate the high-quality image reconstruction and flexibility of VGenNet.
Article
Optics
Hao Wang, Jiabei Zhu, Jangwoon Sung, Guorong Hu, Joseph Greene, Yunzhe LI, Seungbeom Park, Wookrae Kim, Myungjun Lee, Yusin Yang, Lei Tian
Summary: Topography measurement is crucial for surface characterization and inspection applications. This study presents a novel topography technique called Fourier ptychographic topography (FPT), which combines a computational microscope and a phase retrieval algorithm to achieve wide-field-of-view and high-resolution topography reconstruction with nanoscale accuracy. FPT has important implications for surface characterization, semiconductor metrology, and inspection applications.
Article
Nanoscience & Nanotechnology
Jinming Gao, Jinying Guo, Anli Dai, Guohai Situ
Summary: Imaging in scattering media has always been a major challenge due to the interference of background noise with the object information carried by ballistic light, resulting in poor image quality. To address this, active illumination imaging technology offers various advantages over passive imaging by introducing controllable parameters like polarization and coded aperture. In this study, we successfully introduced orbital angular momentum into scattering imaging, effectively enhancing the mid/high frequency components of the object. Furthermore, we fused the enhanced image with the low-quality image obtained from traditional imaging, leading to significant improvements in signal-to-noise ratio and image contrast. This method holds potential applications in foggy environments for autonomous driving, lidar, and machine vision.
Article
Optics
Jianying Hao, Xiao Lin, Yongkun Lin, Mingyong Chen, Ruixian Chen, Guohai Situ, Hideyoshi Horimai, Xiaodi Tan
Summary: In this study, a complex amplitude demodulation method based on deep learning was proposed for holographic data storage. A single-shot diffraction intensity image was used to demodulate both the amplitude and phase by analyzing the correlation between the diffraction intensity features and the encoding data pages. This method achieved multilevel complex amplitude demodulation experimentally without iterations for the first time in HDS.
OPTO-ELECTRONIC ADVANCES
(2023)
Article
Chemistry, Multidisciplinary
Guocui Wang, Jinying Guo, Xinke Wang, Bin Hu, Guohai Situ, Yan Zhang
Summary: This article introduces the methods of using super cells or multi-layer metasurfaces to achieve various multi-functional and exotic functional devices. By mathematically describing the Jones matrix, the computational requirements can be reduced, and devices with various functions can be constructed.
Article
Astronomy & Astrophysics
Situ GuoHai
Summary: Scattering of coherent light and coherent vortex beams propagating through a segment of a multi-mode fiber both produce speckle. Recognizing the orbital angular momentum of the incident vortex beam from the output speckle becomes significantly important for vortex beam study and mode division multiplexing, and deep learning is demonstrated as an efficient approach to address this problem.
SCIENTIA SINICA-PHYSICA MECHANICA & ASTRONOMICA
(2022)
Article
Optics
Zhu Sunyong, Jin Ying, Wu Quanying, Liu Haishan, Situ Guohai
Summary: A hybrid neural network model based on 3D-2D convolution tandem is proposed to overcome the problem of low accuracy in practical flame reconstruction. The model uses 3D convolution to extract spatial features from multiple views and 2D convolution to accelerate training speed. Compared to traditional algorithms, the proposed model achieves higher accuracy with lower time consumption.
ACTA OPTICA SINICA
(2022)
Article
Optics
Junfeng Hou, Guohai Situ
Summary: Optical technologies have been widely used in information security, but the linearity issue in current optical encryption techniques makes the system vulnerable to attacks. Recent research shows that a true nonlinear optical encryption technique can be achieved by utilizing the self-phase modulation effect of photorefractive crystals, providing robustness against known plaintext attacks.