I'm trying to create a mask. I have database of images similar like this image.
CODE
import cv2
import numpy as np
img = cv2.imread('sample1.png', cv2.IMREAD_UNCHANGED)
gray = cv2.cvtColor(img,cv2.COLOR_RGB2GRAY)
#img_ = cv2.threshold(gray,100,225,cv2.THRESH_BINARY)
edges = cv2.Canny(gray, 250, 250)
cv2.imwrite('output.png',edges)
How can I remove inner bonder and fill with white.
Well, there are many ways to do that. All of them need some tuning, depending on your image. There is, for example, a floodfill function in opencv.
But the easiest is probably to use some mathematical morphology and then connected component. Because from the connected component, it is easier to adjust result if needed.
We can start by having a binary version of your edges
binedge=(edges>0).astype(np.uint8)
Once this is done, since there are "holes" in it, we need to fill those holes, so that the edge strictly separate inside from outside. This can be done by a dilatation
ker=np.ones((3,3))
fatedge=cv2.dilate(binedge, ker)
Then, we want to find the inside. That is not easy, because there might be many parts in that inside. So the easiest way is probably to find the outside and revert it. Tho there could also be several outside parts, if character touch the border in different places. So, let's start to find all connected black parts of this picture.
n,comp=cv2.connectedComponents((fatedge==0).astype(np.uint8))
comp here is an image whose values is the index of the connected component. Shown here with random colors for each index.
Let's assume that outside is connected and that (0,0) is in it (it is almost always the case. And it is here. If not, you'll have to find more complex criteria. Such as "the biggest component". Or even to merge different parts). The component we are interested in is the one that contains (0,0). That is the pixels of comp that have the same value as comp[0,0]. And in fact, what we are interested in is the opposite of that: what is inside. We compute outside only because it is easier. Inside is what is not inside, that is pixels that are != comp[0,0].
filled=(comp!=comp[0,0]).astype(np.uint8)
Last stage (not really necessary from an aesthetics point of view. But strictly speaking, it is needed) : since we have dilated the edges at the beginning, this picture is a few pixels bigger than it should. We can erode it back now that we have what we want
output=cv2.erode(filled, ker)*255
cv2.imwrite('output.png',output)
So, all together
import cv2
import numpy as np
img = cv2.imread('Downloads/93Lwd.png', cv2.IMREAD_UNCHANGED)
gray = cv2.cvtColor(img,cv2.COLOR_RGB2GRAY)
#img_ = cv2.threshold(gray,100,225,cv2.THRESH_BINARY)
edges = cv2.Canny(gray, 250, 250)
# Binarize edges
binedge=(edges>0).astype(np.uint8)
# Removing edges too close from left and right borders
binedge[:,:20]=0
binedge[:,-20:]=0
# Fatten them so that there is no hole
ker=np.ones((3,3))
fatedge=cv2.dilate(binedge, ker)
# Find connected black areas
n,comp=cv2.connectedComponents((fatedge==0).astype(np.uint8))
# comp is an image whose each value is the index of the connected component
# Assuming that point (0,0) is in the border, not inside the character, border is where comp is == comp[0,0]
# So character is where it is not
# Or, variant from new image: considering "outside" any part that touches one of the left, right, or top border
# Note: that is redundant with previous 0ing of left and right borders
# Set of all components touching left, right or top border
listOutside=set(comp[:,0]).union(comp[:,-1]).union(comp[0,:])
if 0 in listOutside: listOutside.remove(0) # 0 are the lines, that is what is False in fatedge==0
filled=(~np.isin(comp, list(listOutside))).astype(np.uint8) # isin need array or list, not set
# Just to be extra accurate, since we had dilated edges, with can now erode result
output=cv2.erode(filled, ker)
cv2.imwrite('output.png',output*255)
