Data preparation

raw images from Nikon Ti-E is 2048x2048 (2048x2044)

the data set used by segnet is 480x360 (original camvid dataset)

the dataset used by U net is 696x520

VOC data set used by FCNs (UCBerkeley) is 500x281

Current configuration;512x512 for segnet takes ~7GB of GPU RAM


best results, remove the background in imageJ through remove background with 10 pixels

make mask 


Make ROI in ImageJ
load the image data in imageJ.
use free hand polygon selection tool to select different annotation set
in ROI manager, save first, then more>combine
process binary>create mask (it doesn't maker if the LUT is reversed or not)
use math>MAX to change the label value
save as new files




annotation set
1. singlet cells
2. touching cells
3. 

Other way to annotate image
How to annotate images 
1. JS segment annotation
https://github.com/kyamagu/js-segment-annotator
2. labelme
http://labelme2.csail.mit.edu
3. labelme annotation tool with python with pyqt interface by wkentaro
http://github.com/wkentaro/labelme
4.Ratsnake
http://is-innovation.eu/ratsnake/

5. VGG annotator
http://www.robots.ox.ac.uk/~vgg/software/via/




How to assemble image pair for segmentation for training using lmdb and hdf5

1. lmdb:DIGITS is easy
but limited to 8 bit or RGB png or jpg files

2. hdf5:
https://stackoverflow.com/documentation/caffe/5344/prepare-data-for-training#t=201706122035441079419



Preparation of training data for DeepCell
use 2x2 binning and 1.5X main magnification on Nikon Ti-E microscope with Hamamatsu Orca-Flash V4.0 sCMOS camera.
The image should be 1024x1022. with 0.89um/pixel resolution

1. Open the image in ImageJ.
2. duplicate
3. Gaussian Blur 3 pixels.
4. subtract gaussian blurred image from the raw image using image calculator with 32-bit float on. (This will alleviate some halo artifact and negate some background of dusts in the optical paths.)
5. save as 8bit files.

old


6. Use freehand tool to track the boundary of cells manually (a Tablet will be useful we use a Wacom Cintiq pro 16)
7. Save the ROI set.
8. in ROI manager, combine selection
9. edit>selection>create mask >intracellular mask
10. duplicate
11. use the new one > binary>outline. invert the LUT. save this as feature_0.png
12. subtract feature_0.png from intracellular mask to make sure all the pixels are either 0 or 255. 13. save this as feature_1.png

Now
no making modified images

6. use freehand tool to track the boundary of cells manually
7. save roi set
8. in roi manager, combine selection
9. create a new file with black background with 1024x1022. draw all the selection to 255.
10. save as feature_0.png
11. edit>selection>create mask>intracellular mask
12. duplicate
13. subtract feature_0 from intracellular mask to make sure all the pixels are either 0 or 255.
14. save this as feature_1.png
15. plugins>LOCI>ROI map->save this as labeled.png


Comments