Content Aware Fill for Linux's Python

A simple baseline for image inpainting (not learning-based method, implemented in pure C/C++).

The program will automatically fill in masked region via Content Aware Fill algorithm. An online interactive demo can be found here.

I borrowed from other's repo, and I made it usable for Linux's Python.

Build from Source

You can build it from source by (assume pybind11 is installed by python):

git clone [email protected]:Karbo123/content-aware-fill.git --depth=1
cd content-aware-fill
# build & install
make install-content-aware-fill
# just build, don't install
# make content-aware-fill

Usage

You can use it as follows:

cd python && ipython

# inside python, run:
import libcontent_aware_fill as cwf, cv2
img = cv2.imread("../assets/cat.png")
mask = cv2.imread("../assets/mask.png", 0)
img_masked = img * (~mask[..., None]).astype(bool)
out = cwf.content_aware_fill(img_masked, mask)
cv2.imwrite("../assets/cat_img_masked.png", img_masked)
cv2.imwrite("../assets/cat_mask.png", mask)
cv2.imwrite("../assets/cat_out.png", out)
# result images have been written to:
#   ../assets/cat_img_masked.png  (input)
#   ../assets/cat_mask.png        (input)
#   ../assets/cat_out.png         (output)

API

The full API usage can be printed out:

cd python && ipython

# inside python, run:
import libcontent_aware_fill as cwf
cwf.content_aware_fill?
# Perform Content Aware Fill Algorithm for `image` given the inpainting mask `mask`
# Params:
#  - image (numpy.ndarray): shape=(H, W, C), uint8, row-major, contiguous, C=1(Gray)/2(GrayA)/3(RGB)/4(RGBA)
#  - mask (numpy.ndarray): shape=(H, W), uint8, row-major, contiguous, 255 for regions to be filled
# Return:
#  - image_out (numpy.ndarray): shape=(H, W, C), uint8, C=1(Gray)/2(GrayA)/3(RGB)/4(RGBA)
# Full API:
#     Array content_aware_fill(
#         Array image, Array mask,
#         bool isMakeSeamlesslyTileableHorizontally=false,
#         bool isMakeSeamlesslyTileableVertically=false,
#         int matchContextType=1,
#         float mapWeight=0.5f,
#         float sensitivityToOutliers=0.117f,
#         int patchSize=30,
#         int maxProbeCount=200
#     )

For matchContextType (i.e. type of fitting method), its meaning is as follows:

Context	Method	Shape	Direction
0	Patching	-	-
1	Shuffle	Random	Random
2	Brushfire	Concentric	Inward
3	Directional	Horizontal	Inward
4	Directional	Vertical	Inward
5	Brushfire	Concentric	Outward
6	Directional	Horizontal	Outward
7	Directional	Vertical	Outward
8	Squeeze	Concentric	Both