Thermal Image Segmentation (Part 2)

In the previous topic we presented a simple way for thermal image segmentation which was adapted to simple images. Now, we will try to process a lit bit complicated thermal images such as this one:

This picture comes from one of our research program in hospital, and as it can be seen, the thermal image is really noisy with a lot of temperature variations. In this case, a RGB approach is unadapted and gives very poor results. 

To process such images, the way I adopted is a semantic segmentation approach which involves the use of neural networks to accurately segment images.

After some research, I found a fantastic library, PixelLib, which is written by a talented programmer, Ayoola Olafenwa. Her code is available here: https://github.com/ayoolaolafenwa/PixelLib.

This library is written with Python. But, since Red  allows to call external programs, it was really easy to use the Ayoola's library, and the result is perfect. Basically the redCV code extracts a RGB image from the thermal image and then applies semantic segmentation for each pixel.

Here we use an instance segmentation method and the program identifies the object as a cup with a probability = 0.99. Good performance.

You'll find here https://pixellib.readthedocs.io/en/latest/ all the documentation. Thanks a lot to Ayoola for sharing her code.