Adding Autocropper V2 to main #4

ewellenr · 2023-10-21T19:30:21-04:00

ewellenr commented

2023-10-21 19:30:21 -04:00

Allowing the use of V2 of the autocropper to be added to main for others to use.

ewellenr added 31 commits 2023-10-21 19:30:21 -04:00

Initializing rotator ML algo 61cf15e99c

Pulling the rotator ML from the app branch and putting it in this one,
the correct one.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

First complete implementation of ML Model 96a4abf6aa

Doesn't really work well after training but work in progress.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

State after first attempt at training 71e712e55d

Tried training for a while with current state. Didn't show any
progress with training so I'm trying something else. Might also
just need fewer parameters but we'll see.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

First adjustment to try and get training to work. c724e3be78

Added a batch normalization layer and adjusted
the file to accomodate a restart in the training process.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Merge branch 'main' of ssh://ssh.git.ewellenr.ca:2222/ewellenr/receipt_indexer into autocropper d25dbf576a

End of V1.1 Attempts 38b332f5c9

End of trying to train my own v1.1 custom model for autorotation.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Pushing current training loop and other helpful files c3e9f73e73

saveaspng.ipynb: Saves the dataset as individual .jpg files
	so that bad images can be more quickly seen using OS file
	previews
dataset_viewer.ipynb: Let's you iterate through the individual
	images of a dataset. A less helpful version of what saveaspng
	and OS file previews would do
image_viewer.ipynb: Nearly identical to dataset_viewer.ipynb
manualrotationchecker.ipynb: First version of a manual (non-ML)
	autorotater/deskewer
testcropper.ipynb: New version of a manual autocropper
manualcropandrotate.ipynb: Combining manual cropping and rotating
Also updated the training loop file and added a blacklist for when
making the dataset from the original dataset. Finally, the
dockerfile was updated to remove installation of some unused libraries
and added a library for the manual autorotator.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Point before adding search selection segmentation c921f03df2

About to add search selection segmentation so that
I can have multiple processes since some fail to work
correctly while others do.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Before trying houghline technique 4500f79772

Just a checkpoint

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Start of Hough Line deskew f66a757d8b

Exactly what the title says.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Deskew using Hough Lines d5e7a2eed2

First implementation finished. Working on using it to crop.
Going to try using probabalistic hough lines and a bounding box
to pick out lines within a margin of the correct orientation (vertical)
and put a bounding box around these lines to try and approximate the
receipt.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

First implementation of houghline cropping/deskewing 5bd770da55

Still need to fix threshold/morphology so that the bed
picture works.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Merge branch 'main' of ssh://ssh.git.ewellenr.ca:2222/ewellenr/receipt_indexer into autocropper-test a62f628cc1

First complete implementation of hough line deskewing b2f3e89014

Now to work on hough line cropping.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

First implementation of houghline cropping. 4c7bb2c54f

Need to fix because deskewing still has some issues.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Second iteration of houghline cropping and deskewing. 9c2d801be3

Now have deskewing and cropping as one function.
Need to modify it still so it refines the photo even more
and also returns just the rotated rectangle of the original image
instead of the effected (by morphology) image.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Third checkpoint in making houghline crop and deskew 7f796c7a7d

Currently breaking it down in to more digestable
functions and then need to work on post processing so
that the main piece is kept and the background is set to
white and the main piece of paper is set to completely white
with text being full black.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Cropped original image using functions. 834f562604

Houghline crop now returns a cropped version of
the original image (no threshold/black-and-white or
a shrunken picture (smaller size). Just the base image cropped
to what it should be. Need to now add post processing
to make the background white and sharpen the image (make the text
a hard black and the page a hard white).

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Complete houghline deskewing and cropping. 882892cb53

Now to implement cleaning up the resulting image so that
the the outer area and page is white and the text is black.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

V1 of final postprocessing for houghline 264f8fa80a

Has it working but it isn't tuned so it needs to
be tuned/adjusted so that it works well.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Quick checkpoint with postprocessing. 15b36b2de0

Still struggling with getting the text set well while
removing all the other noise and such.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Quick checkpoint 2 with post processing. 3a98ad0d1c

I believe the idea is to use morphology to clean
up the merged threshold version and then get bounding boxes
for the letters, get a mask from those bounding boxes and
and then apply the merged thresh onto a white page using the mask.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

V2 houghline preprocessing 011acea572

Final version of V2 of houghline preprocessing.
May need to make changes to this version but it's complete
and ready for the OCR and actual ML part now.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Cleaning up and putting functions into python file 518cea9968

As title says, cleaning up and putting the used/important
functions into myfunctions.py file so they could be easily used.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

V2.1 Tuning update b25ffb8602

Just tuning the threshold a bit so that the
background whiteout works better.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Updating gitignore for python function file. eb04834d66

As title says. The python file seems to make
a cache directory so we are ignoring it.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Updated Plan for autocropper 2888e40ee2

Title. Plan is to just convert all the houghline
processing into C since it should be faster than the python opencv.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Merge branch 'autocropper' of ssh://ssh.git.ewellenr.ca:2222/ewellenr/receipt_indexer into autocropper-test e0ce309a0e

Cleanup commit 423b511dd9

Moving around the testing notebooks. Autocropping is about done
with exception to any new versions or converting the stuff to C
code.

Signed-off-by: Ethan Wellenreiter <ewellenreiter@gmail.com>

Merge pull request 'Merging Py V1 of the completed autocropper/cleanup' (#2 ) from autocropper-test into autocropper f91fb1f9b1