A5017194489- Neural Networks and Reservoir Computing
- Optical Network Technologies
- Photonic and Optical Devices
- Digital Holography and Microscopy
- Image Processing Techniques and Applications
- Advanced Fluorescence Microscopy Techniques
- Random lasers and scattering media
- Advanced Vision and Imaging
- Photoacoustic and Ultrasonic Imaging
- Cell Image Analysis Techniques
- Optical Coherence Tomography Applications
- Advanced X-ray Imaging Techniques
- Advanced Photonic Communication Systems
- Sparse and Compressive Sensing Techniques
- Advanced Optical Sensing Technologies
- Advanced Memory and Neural Computing
- Advanced Image Processing Techniques
- Optical measurement and interference techniques
- Power Systems and Renewable Energy
- Advanced Optical Imaging Technologies
- Ear and Head Tumors
- Video Surveillance and Tracking Methods
- AI in cancer detection
- Microfluidic and Bio-sensing Technologies
- Neural dynamics and brain function
Tsinghua University
2014-2024
Beijing Municipal Education Commission
2022
Shenzhen Bao'an District People's Hospital
2021
University of California, Los Angeles
2018-2019
California NanoSystems Institute
2018-2019
Bioengineering Center
2019
South China University of Technology
2018
Northeastern University
2018
Stanford University
2016
Human Media
2013
We introduce an all-optical Diffractive Deep Neural Network (D2NN) architecture that can learn to implement various functions after deep learning-based design of passive diffractive layers work collectively. experimentally demonstrated the success this framework by creating 3D-printed D2NNs learned handwritten digit classification and function imaging lens at terahertz spectrum. With existing plethora 3D-printing other lithographic fabrication methods as well spatial-light-modulators,...
Holography encodes the three dimensional (3D) information of a sample in form an intensity-only recording. However, to decode original image from its hologram(s), auto-focusing and phase-recovery are needed, which general cumbersome time-consuming digitally perform. Here we demonstrate convolutional neural network (CNN) based approach that simultaneously performs significantly extend depth-of-field (DOF) holographic reconstruction. For this, CNN is trained by using pairs randomly de-focused...
In this Letter we propose the Fourier-space diffractive deep neural network (F-D^{2}NN) for all-optical image processing that performs advanced computer vision tasks at speed of light. The F-D^{2}NN is achieved by placing extremely compact modulation layers Fourier plane or both and imaging planes an optical system, where nonlinearity introduced from ferroelectric thin films. We demonstrated can be trained with learning algorithms saliency detection high-accuracy object classification.
This paper proposes a novel compressive hyperspectral (HS) imaging approach that allows for high-resolution HS images to be captured in single image. The proposed architecture comprises three key components: spatial-spectral encoded optical camera design, over-complete dictionary learning and sparse-constraint computational reconstruction. Our sampling scheme provides higher degree of randomness the measured projections than previous approaches; robust nonlinear sparse reconstruction method...
Multispectral cameras collect image data with a greater number of spectral channels than traditional trichromatic sensors, thus providing information at higher level detail. Such are useful in various fields, such as remote sensing, materials science, biophotonics, and environmental monitoring. The massive scale multispectral data-at high resolutions the spectral, spatial, temporal dimensions-has long presented major challenge spectrometer design. With recent developments sampling theory,...
Training an artificial neural network with backpropagation algorithms to perform advanced machine learning tasks requires extensive computational process. This paper proposes implement the algorithm optically for in situ training of both linear and nonlinear diffractive optical networks, which enables acceleration speed improvement energy efficiency on core computing modules. We demonstrate that gradient a loss function respect weights layers can be accurately calculated by measuring forward...
This Letter presents a new snapshot approach to hyperspectral imaging via dual-optical coding and compressive computational reconstruction. We demonstrate that two high-speed spatial light modulators, located conjugate the image spectral plane, respectively, can code datacube into single sensor such high-resolution signal be recovered in postprocessing. show various applications by designing different optical modulation functions, including programmable spatially varying color filtering,...
This paper proposes a novel approach for high-resolution light field microscopy imaging by using camera array. In this approach, we apply two-stage relay system expanding the aperture plane of microscope into size an lens array, and utilize sensor array acquiring different sub-apertures images formed corresponding lenses. By combining rectified synchronized from 5 × viewpoints with our prototype system, successfully recovered color videos various fast-moving microscopic specimens spatial...
Abstract The development of deep learning and open access to a substantial collection imaging data together provide potential solution for computational image transformation, which is gradually changing the landscape optical biomedical research. However, current implementations usually operate in supervised manner, their reliance on laborious error-prone annotation procedures remains barrier more general applicability. Here, we propose an unsupervised transformation facilitate utilization...
Photonic neural networks perform brain-inspired computations using photons instead of electrons to achieve substantially improved computing performance. However, existing architectures can only handle data with regular structures but fail generalize graph-structured beyond Euclidean space. Here, we propose the diffractive graph network (DGNN), an all-optical representation learning architecture based on photonic units (DPUs) and on-chip optical devices address this limitation. Specifically,...
The rapid development of artificial intelligence (AI) facilitates various applications from all areas but also poses great challenges in its hardware implementation terms speed and energy because the explosive growth data. Optical computing provides a distinctive perspective to address this bottleneck by harnessing unique properties photons including broad bandwidth, low latency, high efficiency. In review, we introduce latest developments optical for different AI models, feedforward neural...
Abstract Photonic neural networks are brain-inspired information processing technology using photons instead of electrons to perform artificial intelligence (AI) tasks. However, existing architectures designed for a single task but fail multiplex different tasks in parallel within monolithic system due the competition that deteriorates model performance. This paper proposes novel optical multitask learning by designing multiwavelength diffractive deep (D 2 NNs) with joint optimization...
Following the explosive growth of global data, there is an ever-increasing demand for high-throughput processing in image transmission systems. However, existing methods mainly rely on electronic circuits, which severely limits throughput. Here, we propose end-to-end all-optical variational autoencoder, named photonic encoder-decoder (PED), maps physical system into optical generative neural network. By modeling noises as variation latent space, PED establishes a large-scale unsupervised...
Conventional bioaerosol sensing requires the sampled aerosols in field to be transferred a laboratory for manual inspection, which can rather costly and slow, also requiring professional labeling microscopic examination of samples. Here we demonstrate label-free using field-portable cost-effective device based on holographic microscopy deep-learning, screens bioaerosols at throughput 13 L/min. Two different deep neural networks are designed rapidly reconstruct amplitude phase images captured...
The diffractive deep neural network (D2NN) has demonstrated its importance in performing various all-optical machine learning tasks, e.g., classification, segmentation, etc. However, deeper D2NNs that provide higher inference complexity are more difficult to train due the problem of gradient vanishing. We introduce residual (Res-D2NN), which enables us substantially networks by constructing blocks learn mapping functions. Unlike existing plain D2NNs, Res-D2NNs contribute design a learnable...
Abstract Quantitative volumetric fluorescence imaging at high speed across a long term is vital to understand various cellular and subcellular behaviors in living organisms. Light-field microscopy provides compact computational solution by the entire volume tomographic way, while facing severe degradation scattering tissue or densely-labelled samples. To address this problem, we propose an incoherent multiscale model complete space for quantitative 3D reconstruction complicated environments,...
Abstract Early detection and appropriate medical treatment are of great use for ear disease. However, a new diagnostic strategy is necessary the absence experts relatively low accuracy, in which deep learning plays an important role. This paper puts forward mechanic model uses abundant otoscope image data gained clinical cases to achieve automatic diagnosis diseases real time. A total 20,542 endoscopic images were employed train nine common convolution neural networks. According...
Abstract The comprehensive analysis of biological specimens brings about the demand for capturing spatial, temporal and spectral dimensions visual information together. However, such high-dimensional video acquisition faces major challenges in developing large data throughput effective multiplexing techniques. Here, we report snapshot hyperspectral volumetric microscopy that computationally reconstructs profiles high-resolution volumes ~1000 μm × 1000 500 at rate by a novel four-dimensional...
Abstract Wireless sensing of the wave propagation direction from radio sources lays foundation for communication, radar, navigation, etc. However, existing signal processing paradigm arrival estimation requires frequency electronic circuit to demodulate and sample multichannel baseband signals followed by a complicated computing process, which places fundamental limit on its speed energy efficiency. Here, we propose super-resolution diffractive neural networks (S-DNN) process electromagnetic...
The ability of camera arrays to efficiently capture higher space-bandwidth product than single cameras has led various multiscale and hybrid systems. These systems play vital roles in computational photography, including light field imaging, 360 VR camera, gigapixel videography, etc. One the critical tasks imaging is matching fusing cross-resolution images from different under perspective parallax. In this paper, we investigate reference-based super-resolution (RefSR) problem associated with...