Table of Contents

Problem

Below is a sketch that Siqi made on the whiteboard about the setup. Here is my understanding of the problem (I'm not an expert in the field).

• The beam goes through the DMD
• it is then split
  • first view on the VCC - virtual cathode ?
  • second, after going through nonlinear cathode and its optics, on the YAG

The end goal is to shape and control the beam. To do this one needs to know what happens to the beam as it goes through the cathode.

The cathode creates a non linear gain map - some parts magnify, and some parts shrink. The problem is basically calibration, figuring out what the cathode does to the beam. You can't just let the whole beam through, you need to let a little bit of the beam through at a time and see what the cathode does by looking at the YAG.

• operators can control how big of a square to open up on the DMD to let beam go through.
• A scan across the DMD is just going row by row, opening a different square of DMD pixels to let beam go through.
• One want to measure two things
  • The location of the beam on the VCC
  • The charge of the beam on the YAG
• The charge will be the sum inside the box around the beem on the YAG, so you need to know location on YAG for this

For a given file, to accurately compute charge

• first subtract the background (upper row of above plot)
• Then figure out where the beam is for YAG (bottom row, when all DMD open to let beam through)
• when you sum charge in the YAG box, restrict to the region on the YAG open beam plot, 

Image Added

Screen Information

File 4 also contains the yag and vcc backgrounds, as well as shots where the entire beam is opened up (as opposed to just one location on the DMD). Here are those plots

Image Added

Notice the distortion on the yag. This is because the beam goes through the cathode and its nonlinear optics before showing up on the yag screen. 

Data

The page Accelerator Beam finding - Internal Notes talks about the data and code, page has restricted access.

...

• this is a vcc screen, and the location of the beam has been labeled with the white box. 
• We want to use machine learning to predict these box locations. 
• The dataset contains vcc screens and YAG screens - two different datasets, currently looking at training two different models - a YAG predictor and a VCC predictor.

Description

There are 3 files, called 1,2 and 4.

• Files 1 and 2 have 142 samples. With file 4, the total number of 239 samples is.
• Each sample has a yag, vcc, and box for each - there are also backgrounds to subtract for file 4, and a oval like image of the entire beam region - one can use that  to narrow the search so as to not predict a box in a corner of the yag or vcc screen
  • File 4 also containts the bkg and beam shots above.
  • Files 1 and 2 already have the bkg subtracted.
• vcc values are in [0,255], and the boxed beam can get quite brite
• yag values go over 1000, I think, but the boxed value is always dim, like up to 14

...

However with the yag images, there is very little difference between nobeam and beam:

There a

At first I suspect thought we will would not be able to do much with these the yag screen codewords without more preprocessing. This may be the case for the classification problem of the yag images - I think they are too faint for what vgg16 expects - it was trained on the imagenet color imageswether or not the beam is present, but for the regression problems of find the box around the beam, assuming it is there, it actually does better on yag than vcc.

...

Preprocessing

This problem seems harder than the localization for lasing fingers in amo86815. There is more variety in the signal we are trying to find . This leads to different kinds of signal processing pre-filtering of the images. Then sometimes the vgg16 codewords don't seem that homogenous - suggesting.on the vcc. 

Of the 239 samples, 163 of the vcc have a labeled box. Below is a plot where we grab what is inside each box and plot it all in a grid - this is with the background subtraction for file 4. The plot on the left is before, and on the right, is after reducing the 480 x 640 vcc images to (224,224) for vgg16. We used scipy imreduce 'lanczos' to reduce (this calls PIL). Here, there is no preprocessing other than what the image size reduction does

Here are the 159 smaples of the yag with a box - here are are using 'lanczos' to reduce from the much larger size of 1040 x 1392 to (224,224). It is interesting to note how the colorbar changes - the range no longer goes up to 320 - I think the 320 values were isolated pixels that get washed out? Or else there is something else I don't understand - we are doing nothing more than scipy.misc.imresize(img,(224,224), interp='lanczos',mode='F') but img is np.uint16 after careful background subtraction  - (going through float32, thresholding at 0 before converting back)

...

For the yag, a 1% overlap is 86%:

All Accuracies

36 different runs were carried out, varying each of the following:

• Pre-processing algorithm, one of
  • none
    • just 'lanczos' reduction
  • denoise-log
    • 3 pt median filter
    • log(1+img)
    • 'lanczoz' reduction
    • multiply by scale factor
  • denoise-max-log
    • 3 pt median filter
    • 3 x 3 sum
    • 3 pt median filter
    • 'max_reduce' (save largest pixel value over square)
    • 'lanczoz' reduction (to get final (224,224) size)
    • log(1+img)
    • scale up
• files, one of
  • just 1,2
  • 1,2,4
• Do and Don't subtract background for file 4
• Do and Don't filter out some of the 8192 features with variance <= 0.01 before doing regression

Below is a table of all these results

nm=yag eb_alg_none_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.41 th=0.20 acc=0.78 th=0.01 acc=0.86
nm=vcc eb_alg_none_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.14 th=0.20 acc=0.38 th=0.01 acc=0.65
nm=yag eb_alg_denoise-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.66 th=0.20 acc=0.87 th=0.01 acc=0.90
nm=vcc eb_alg_denoise-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.38 th=0.20 acc=0.61 th=0.01 acc=0.72
nm=yag eb_alg_denoise-max-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.46 th=0.20 acc=0.77 th=0.01 acc=0.88
nm=vcc eb_alg_denoise-max-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.41 th=0.20 acc=0.60 th=0.01 acc=0.76
nm=yag eb_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.34 th=0.20 acc=0.68 th=0.01 acc=0.82
nm=vcc eb_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.11 th=0.20 acc=0.34 th=0.01 acc=0.61
nm=yag eb_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.47 th=0.20 acc=0.75 th=0.01 acc=0.89
nm=vcc eb_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.26 th=0.20 acc=0.42 th=0.01 acc=0.56
nm=yag eb_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.34 th=0.20 acc=0.62 th=0.01 acc=0.81
nm=vcc eb_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.28 th=0.20 acc=0.39 th=0.01 acc=0.48
nm=yag eb_subbkg_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.38 th=0.20 acc=0.73 th=0.01 acc=0.86
nm=vcc eb_subbkg_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.08 th=0.20 acc=0.28 th=0.01 acc=0.55
nm=yag eb_subbkg_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.56 th=0.20 acc=0.81 th=0.01 acc=0.92
nm=vcc eb_subbkg_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.23 th=0.20 acc=0.46 th=0.01 acc=0.65
nm=yag eb_subbkg_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.27 th=0.20 acc=0.64 th=0.01 acc=0.84
nm=vcc eb_subbkg_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.26 th=0.20 acc=0.48 th=0.01 acc=0.63
nm=yag eb_varthresh_alg_none_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.41 th=0.20 acc=0.79 th=0.01 acc=0.86
nm=vcc eb_varthresh_alg_none_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.10 th=0.20 acc=0.38 th=0.01 acc=0.63
nm=yag eb_varthresh_alg_denoise-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.67 th=0.20 acc=0.87 th=0.01 acc=0.90
nm=vcc eb_varthresh_alg_denoise-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.37 th=0.20 acc=0.58 th=0.01 acc=0.71
nm=yag eb_varthresh_alg_denoise-max-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.45 th=0.20 acc=0.77 th=0.01 acc=0.88
nm=vcc eb_varthresh_alg_denoise-max-log_f1_f2-regress.h5 inter/union accuracies:  th=0.50 acc=0.40 th=0.20 acc=0.59 th=0.01 acc=0.76
nm=yag eb_varthresh_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.30 th=0.20 acc=0.67 th=0.01 acc=0.82
nm=vcc eb_varthresh_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.10 th=0.20 acc=0.33 th=0.01 acc=0.60
nm=yag eb_varthresh_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.47 th=0.20 acc=0.75 th=0.01 acc=0.89
nm=vcc eb_varthresh_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.24 th=0.20 acc=0.42 th=0.01 acc=0.57
nm=yag eb_varthresh_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.34 th=0.20 acc=0.62 th=0.01 acc=0.81
nm=vcc eb_varthresh_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.28 th=0.20 acc=0.39 th=0.01 acc=0.48
nm=yag eb_varthresh_subbkg_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.35 th=0.20 acc=0.72 th=0.01 acc=0.86
nm=vcc eb_varthresh_subbkg_alg_none_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.07 th=0.20 acc=0.25 th=0.01 acc=0.53
nm=yag eb_varthresh_subbkg_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.54 th=0.20 acc=0.79 th=0.01 acc=0.92
nm=vcc eb_varthresh_subbkg_alg_denoise-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.22 th=0.20 acc=0.44 th=0.01 acc=0.64
nm=yag eb_varthresh_subbkg_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.27 th=0.20 acc=0.65 th=0.01 acc=0.84
nm=vcc eb_varthresh_subbkg_alg_denoise-max-log_f1_f2_f4-regress.h5 inter/union accuracies:  th=0.50 acc=0.26 th=0.20 acc=0.47 th=0.01 acc=0.62

Best Result - YAG

The best 1% overlap for the YAG is 92%

...

Best Results 

The best results have been obtained using some signal preprocessing developed by Adbullah Ahmed. The de-noising is roughly:

• vcc: threshold at 255

...

Image Removed

Best Result - VCC

The best 1% overlap for the VCC is 76%. 

• It is over files 1,2
• used denoise-log
• adding file 4, with subbkg, reduced acc to 63%
• adding file 4, without subbkg reduced acc to 48%
  

Image Removed

Third Pass 

...

• opencv medianBlur
  • yag: 5pt
  • vcc: 7pt
• opencv guassianBlur
  • yag: 55 x 55
  • vcc: 15 x 15
• yag: threshold at 1.5 (where < 1.5, set to 1.5) 
• lanczos reduction

After doing the de-noising, and before the reduction, we find the maximum value in the image and call it a hit if it is in the labeled box. This signal processing solution performs quite well. Over files 1,2,4 and doing the background subtraction for file 4, it does:

...

• yag: inter/union accuracies: th=0.50 acc=0.89 th=0.20 acc=0.97 th=0.01 acc=0.98

• vcc: inter/union accuracies: th=0.50 acc=0.09 th=0.20 acc=0.38 th=0.01 acc=0.66

Preprocessing Plot

Here is a plot of the preprocessing - the main benefit from it is the de-noising, which is not as apparent in this plot

Image Added

Image Added

Here are plots to show the regression results

vcc median+Gaussian Blur

yag median+Guassian Blur