- I have used basNET for differentiating the background from the foreground.
- Background pixels are closer to [255,255,255] and foreground are closer to [0,0,0] using this property I converted all the pixels based on their Euler distances from these pixels.
- Then used these pixels values completely remove the foreground image so the mean shift algorithm can be applied only on the background image.
- I used the sklearn.cluster libraries mean shift algorithm. I got a result that looks reasonable. But the time it is taking is too much. Need to search for some better implementation.
- Regarding searching for the maximized space, I was thinking of using DFS to calculate the number of pixels that have approximately the same pixel values.
- Once we get values of the region for all components we can take one with maximum area.
- One noticeable problem in this approach is that the region with the maximum area may not be in rectangular or some reasonable shape.
- So if there is some built-in library to do this, it will be easy. I need to see if there exists one.
These are some results of the model.
Use the code mentioned in Imageprocessing.ipynb to execute. Run the cells one after the other. As of now I have used the code for individal part from other resources. But we will specialize them very soon.