SRGS 2022

General Info

• Potential Projects Summaries
• Sharepoint Link

AI Feature Recognition: Extract Spectrometer Angle from Image

Students:

• Anna Rosner
• William Savage

Useful links/info:

• angle-cam-image-recognition.pdf
• Location of example images: /work/hallc/shms/spring17_angle_snaps/
  • Time the image was acquired is embedded in the image file
  • The numbers in the snapshot filenames are the run numbers
  • 4,265 images ; ~92kB/file ; 391MB total
• The value of the encoders are stored in the MYA EPICS archive
  • PV names are:
    • ecSHMS_Angle
    • ecHMS_Angle

Initial thoughts from Brad

   I had been imagining splitting the photos into two regions: one with
   the digits, and a second with the vernier scale. Each region would be
   evaluated/interpreted separately with some 'optimized' algorithms.
   
   'Real' errors/discrepancies would be best indicated by a scanning for a
   mismatch between MYA and the analysis database record and/or the value
   flagged in the logbook which has generally been vetted and updated by a
   human. The simplest way to test 'bad' angles would be just to
   (randomly) shift the truth angle by a small amount -- that would be
   indistinguishable from an observed drift in the EPICS encoder system.
   
   I (or the students) can also look for angle shifts in the 'real' data,
   but that will take some poking around. It should be indicated by a
   sharp (small) jump in the MYA value as an offset is changed to bring
   the EPICS value in agreement with the camera readback.
   
   One other dataset that I could obtain is a movie of the angle changing
   over a range (the movie is just a compilation of frame grabs). The
   individual frames could be pulled out of the mp4 and evaluated
   individually over a continuously varying range of angles.