I have read that one of the requirement of image inpainting is to have a mask which locates places of an image where inpainting is needed. How is this mask automatically designed? From the internet I searched but I can't get an explicit information on how to automatically locate scratches in an image (mask). I also read that one can locate those damaged parts by taking the difference of the damaged image and the median-filtered original image. This doesn't give me better results, in fact I get extremely bad results. If one has a MATLAB function, please share.
It depends on the reason why you are doing inpainting:
- if you want to remove an object from the scene then either you manually delineate that object or you run some image analysis algorithm to find it
- if you want to inpaint some type of damage (cracks in a painting, missing blocks of a video stream) then again either you manually specify the holemap or you need an algorithm that can detect damaged regions
If you have the original image just look at the pixels where your scratched image is different from the original. Obviously in real applications you will not have the luxury of having this true original and this method is only useful for testing your inpainting algorithm on synthetic examples.
Automatic crack detection is not a trivial task.
Thanks very much for the valuable comment. However, I find it impractical or relatively challenging to manually search for the cracks by analyzing the pixel values in the image. My greatest challenge is to create a mask (even the manual process). Assume that the objective is to remove cracks in the image such as the one attached with this comment. How is the Mask created for this image? What are the steps needed. In the comment given by Beerend, did I get the comment right that I have to manually use the normal photo editing tools to analyze the pixels.
Like I said, automatically detecting cracks is a challenge on its own. It is hard to differentiate between an edge in the actual image or a discontinuity due to physical damage.
Without automated crack detection you still have the following options options:
- you manually specify the entire mask, which is indeed a tedious and boring task
- you manually specify some pixels of the crack and use a region growing algorithm to automatically find the rest of it
- this one is a bit 'cheating' but you could also use someone else's inpainting result and subtract the damaged original image, the nonzero pixels should give you the inpainting mask
- Badges
- Science topic