Friday, May 03, 2019 (put topic here)¶
Attendants: Ulf, Richard, Malin, Luana, Maren, Josefine, Katharina, Sophia, Thomas, Michael
Summary of last week’s work¶
Classification decision tree:¶
Firs decision tree is created:
“Hollow” is missing.
Research and Library¶
New papers have been added to the repository.
Preprocessing and naming conventions¶
Cut images, give each one a label
Naming convention (name, ID, date, time)
Put each peace in a new folder with: id_number_specific number
Input of image preprocessing: e.g. G01-190411-092055-873_F00.bmp (contains date, batch number and position on band)
Output of image preprocessing: e.g. 0_0.jpg = id_number, where ‘id’ is the newly assigned ID and ‘number’ takes on values from 0 to 2, indicating the position on the band. The mapping of ID to original name is maintained, because the preprocessor creates a csv file that stores for each id what the original name was.
Output of image augmentation: e.g. rot2_0_0.jpg = pre#_id_number, where ‘pre’ is a 3-letter code to identify the transformation that was applied and # is the index (e.g. rot=rotation creates 3 images, so rot2 would be the third one of these, where the image was rotated to the right.)
Transformation codes:
rot = rotation
mir = mirroring (along y axis)
tra = translation (by some random number of pixels)
This process is recursive. Transformations are applied in layers, so the output of one augmentation is the input for the next augmentation. e.g. tra1_mir0_rot0_3_1.jpg is id=3, second image, first rotation, first mirroring, second translation. Rotation was applied first, translation last.
We can create a lot of labelled artificial data from a small source.
Integrate preprocessing into the pipeline (Ulf)
Think of a database/table to store the characteristics of a Pic-ID.
Label software¶
Give the created images a label by adding a prefix.
We should continue to build an interface that makes it easy for us to label data ourselves.
Service¶
The service was tested on the machine. However, the machine will output an error because it is trying to access the directly moving images.
It is possible to start and stop the service by pressing the Windows key + enter “Service” and open the window. Search for Directory Check and use the GUI.
Team Viewer¶
plan to manually collect and move images using the Google Calendar.
During the meeting¶
We had a discussion about how to use git. It will be finished.