A lightwieght image inpainting tool written in python. Simple and effective tool to remove scratches, bruises and small holes in images. Could be easily integrated at the backend for Flask or Django apps related to image restoration.

Inpainting is a process of image restoration, the idea is to fill the damage, deteriorated or missing part of an image to present a complete image. It can also be used to remove unwanted parts in an image. Deep learning based approaches use GANs to inpaint. This requires significant amount of training. The proposed tool quickly inpaints by solving a PDE on graphs.

pip install pyinpaint

• Command line

pyinpaint --org_img "path/to/original/image" --mask "path/to/mask"
# pyinpaint --org_img  --mask  --ps --k_boundary --k_search --k_patch

The output is an inpainted image at the path of org_img.

• Python

from pyinpaint import Inpaint
inpaint = Inpaint(org_img, mask)
inpainted_img = inpaint()
#inpaint = Inpaint(org_img, mask, ps)
#inpainted_img = inpaint(k_boundary, k_search, k_patch)

This returns a numpy array inpainted_img.

This video shows:

• How to create an inpainting mask in Gimp.
• How to inpaint using command line.

Basically the inpainting is achieved using harmonic extension on a non-local graph created using image to be inpainted.

• f(u) is the signal (rgb values) on the node u.
• \Delta_{w,p} is the weighted graph p Laplacian.
• A is the area to be inpainted.
• dA is the area where signal is given as g(u).

For p=2 the solution to the above equation yields non-local means.

• w(u,v) is the weight on the edge from node u to v.
• N(u) is the set of neighbors of node u.
• d(u) is the degree at node u.

The following description of the parameters is useful to gain speed-ups and to inpaint low spatial frequency texture images.