PFEE - GE Healthcare - Inpainting

Authors

• Victor Guichard
• Guillaume Valente
• Mehdi Bekhtaoui
• Alexandre Yvart

Content

Data

2D Generators:

from inpainting.data import CustomGenerator, MaskGenerator

2D+t Patch Generation

from inpainting.data import generate_patches

2D+t Mask Generation (creates a volume)

from inpainting.data import Generate3DArtifacts

Layers

This package contains the implementation of Partial Convolution 2D and Partial Convolution 3D

from inpainting.layers import PConv2D, PConv3D

These layers are similar to Conv2D, Conv3D but take as input the image and the mask:

PConv3D(filters, kernel_size, padding='same')([img, mask])
PConv2D(filters, kernel_size, padding='same')([img, mask])

Models

A pre defined model for PConv2D and PConv3D (many2many and many2one) are implemented:

from inpainting.model import PConvModel, PConvModel3D

model = PConvModel(...) # Pre compiled and ready to use 2D model
model = PConvModel3D(...) # Pre compiled and ready to use 3D many2many model
model = PConvModel3D(..., many2many=False) # Pre compiled and ready to use 3D many2one model

Losses

The custom vgg loss are defined inside the models

Data

• Custom Generator: Custom Keras DataGenerator that will output image mask and ground truth
• Mask Generator: Generator that build 2D masks

Contact

If some parts are unclear you can contact: Victor Guichard