I am trying to find, which algorithm or computer vision task(Deep learning task) can achieve following:

My Source Image is:

I want to create segment like:

What type of task or algorithm or series of steps can produce this?

I have tried:

• Segmentation model using Deep Learning. but it does not yield best result always.

I am thinking:

• If we can have combination of OpenCV pre/post processing type of task couples with Deep Learning based semantic segmentations, we can achieve this.

Any suggestions?

CodePudding user response：

This is (semantic) segmentation task in Computer Vision. Deep Learning can be used to do semantic segmentation. There are many methods in deep learning.

You are trying to segment residential area in aerial images as your residential area is white and roads are black in your output mask. But people generally do it reverse i.e. they segment roads. You can find a lot of tutorials (example) on internet by searching "road segmentation in aerial images" . Once you have segmented roads, you can take negative of the output to get black roads.

For best results, you will need labelled data. A quick way would be to use someone else's data (and/or model) and then fine-tune on your own labelled data. You can find other's data on internet (e.g.: Toronto Univ data). You may need around 200-300 of your own labelled images for fine-tuning (transfer-learning).