Unet from based on and Inception like architecture
A network with traditional UNET replaced with Inception like blocks with dilations
Trained using
- Generated images, which have 4 quarters from different images
- The 4 quarters have slight rotation
- All images have 2 channels which give the coordinates in x and y for image
- Augmentation with random flips, random brightness contrast, hue shifts and inversions
Metrics
Mean IOUs
| Body Text | Heading | Floating Text | Caption | Header & Footer | Table | Graphic | Total |
|---|---|---|---|---|---|---|---|
| 93.68 | 62.681 | 55.079 | 69.939 | 90.558 | 73.350 | 68.598 | 74.215 |
Results
- a good increase of about 10 iou from previous train
- drop caps are now treated as part of text. earlier they used to be classified as image
- shadows are now not recognized as image
- better performance with non white pages
Color
body text - yellow
heading - pink
floating - purple
caption - green
header&footer - red
table - dark_blue
graphic - light blue
Images
You can see all the validation set images in the Results folder
