Shanghai Optics Institute and others made progress in intelligent computing imaging research

[ Instrument R & D of instrumentation network ] Recently, the Information Optics and Optoelectronics Technology Laboratory of Shanghai Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, in cooperation with the Institute of Applied Optics, University of Stuttgart, Germany, and the Massachusetts Institute of Technology, proposed and experimentally verified a physical model and The new computational imaging method of deep neural networks does not require a large amount of labeled data to complete neural network training, which will effectively promote the wide application of artificial intelligence technology in computational imaging.
In recent years, deep learning-based methods have been widely used in computational imaging, and have achieved a series of remarkable results in many fields such as phase recovery, digital holography, single-pixel imaging, and scattering imaging. However, traditional computational imaging methods based on deep learning mostly use supervised learning strategies, so it is necessary to obtain a large amount of labeled data in advance to train the neural network, and the quantity and quality of the acquired data have a great impact on the performance of the resulting model. In practical applications, this condition is often difficult to meet. Although previous studies have shown that training data can be generated by simulation when the forward physical model of the imaging system is known, the generalization of neural networks is always limited, and the resulting model can only get good results for scenes similar to the training set. .
Imaging systems are roughly divided into:
â‘ Optical mechanical scanning. Such as multi-spectral scanner. Most use mirrors to scan the object surface, and output image data after spectroscopic, wave detection and photoelectric conversion.
â‘¡Electronic scanning. For example, the back-beam light guide TV camera belongs to the image plane scanning method. The process is optical imaging on the target surface of the light pipe, and the signal is amplified and output after being scanned by the electron beam.
â‘¢Solid self-scanning. For example, the photoelectric scanning sensor of the French SPOT satellite also belongs to the image scanning method. The scene is imaged by an objective lens on a detector array composed of many charge-coupled devices (CCD), and output after photoelectric conversion.
â‘£ Antenna scanning. Such as side-view radar, which is an active remote sensing imaging system that belongs to the object-surface scanning method. It transmits the microwave beam through the antenna and receives the echo reflected by the scene after demodulation and output.
In order to solve the problems of difficulty in obtaining training data and limited model generalization in computational imaging methods based on deep learning, researchers have proposed a method of combining physical models with neural networks, using physical models to replace training data to drive the optimization of network parameters. Compared with the traditional data-driven end-to-end deep learning method, PhysenNet does not need to obtain training data and is a universal method. Compared with the model-driven optimization algorithm, PhysenNet can be used to solve the ill-conditioned inverse problem (recovering the original object information from the detected physical measurement without the loss of information such as phase during the detection phase) without using explicit regular terms.
Compared with the photography system, the advantages of the scanning imaging system are:
① The working band is about 0.38 ～ 14.0 microns, the range is large, and the number of band divisions and the bandwidth of the band can be determined flexibly.
â‘¡Adopt part of the light in the instrument, which is conducive to the accurate registration of images in different wave bands.
â‘¢ The image density after radiation calibration is convenient for machine-assisted processing and classification.
The researchers used a classic example in computational imaging, phase imaging, to verify the effectiveness of the method. Through continuous iterations, the calculated intensity of the diffraction output of the neural network output through the diffraction propagation and measurement process (physical model) gradually approached the actual measurement. As the iteration progresses, the output of the neural network gradually approaches the actual phase object. The experimental results show that when using only a single diffraction intensity map, PhysenNet's recovery effect is superior to that of the Gerchberg-Saxton (GS) algorithm that needs to iterate back and forth between multiple defocus planes, and is close to the recovery of digital holography effect. This method can be applied to many computational imaging methods with known forward physical models.

Our company produce and supply all kinds of original spare parts for water trucks, road sweeper trucks, fire fighting trucks, oil tanker trucks,sewage suction trucks, garbage trucks, roll off garbage trucks, compactor rear loaders..

We sell spare parts contain : PTO , water pump , fuel pump,water cannon, vacuum pump, spraying nozzles, ball valves , flange, seal, sealing ring, hydraulic cylinders, filter, drive shaft, music speakers, liquid level meter, high-pressure pipe, all types rubbish bins...

Water Trucks Spare Parts

water pump, high pressure pump, spray nozzles, PTO, water pipe, fire hose couping, ball valves,water cannon, spray machine

CHENGLI SPECIAL AUTOMOBILE CO.,LTD , https://www.clwcntruck.com