Project: Visible/Near Infrared Image Fusion
Image Fusion Methods (as according to the IEEE paper on Adaptive VS/NIR Fusion Methods):
Blind Approach: Make a direct comparison between visible light image and infrared image and blend the two by mixing (e.g. replacing one channel of the visible light image with the entire NIR image [single channel])
Pros: Quick, Very Simple and Efficient to Compute, Cons: Result has high variance with NIR Image
2. Adaptive Approach: Compute a "Fusion Map" which is an estimate of which parts of the NIR image should be merged with the visible light image
Pros: Result quality does not vary as much with NIR image quality Cons: More compute power required but should still perform on embedded applications
How to get started making the Fusion Map:
Image Derivatives
Learned how to treat a 2D derivative of an image as an Affine kernel convolution, in particular, one that is centered around the pixel of concern (applies no shift on the image), much help from this lecture I found on youtube: https://www.youtube.com/watch?v=1THuCOKNn6U&list=PLd3hlSJsX_Imk_BPmB_H3AQjFKZS9XgZm&index=2
From the Wiki on Image Derivatives, I learned that I could apply it much like we did in the Machine Learning projects as a convolution, a kernel sliding across the picture, though without the right image transformation, it is technically a correlation instead of a convolution
Fusion Map calculation coming soon
Commenti