Dortmund 2021 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AKBP: Arbeitskreis Beschleunigerphysik

AKBP 8: Diagnostics, Control and Instrumentation II

AKBP 8.1: Vortrag

Mittwoch, 17. März 2021, 16:30–16:45, AKBPa

The application of deep learn on slit scan image processing and emittance predict — •Shuai Ma, André Arnold, Anton Ryzhov, Jana Schaber, Jochen Teichert, and Rong Xiang — Institute of Radiation Physics, HZDR, 01328 Dresden, Germany

For slit scan method, how to decrease the noise of beam-let images directly influences the accuracy of the emittance results. There are two kind noise in the images, random noise and dark current. The traditional method is to capture two groups of images, one with beam and the other one without beam as background. The images with beam subtract the background image respectively. Then using filter algorithm, such as Median filter and Gaussian filter, to decrease the random noise. The total time of these is usually 5 to 10 minutes and sometimes the images with beam at the beginning and ending are not very clear because of low signal ratio, which will contribute emittance to the results. To compress the processing time and improve accuracy, one deep learning method, sparse auto-encoder network is used to pre-process the images. To train the network, the slit scan simulation program based on Astra is built to create the image cases. The sparse auto-encoder network is used to filter random noise. During the training, the noise from the real images, background images, is added to increase the stability of the network. After the network, the negative signals, meaningless signals, in the images are set to zero. The other model, point cloud network is used to filter the dark current and gives the emittance from phase space directly. The error is lower than 10%.