I am working on a problem where I have to train a CNN to recognize different kinds of surfaces. One important characteristic of the surfaces I am interested is is how reflective they are. I have been trying to find a method that quantifies how "shiny" a surface is, but I have not found much. I am hoping that someone can point me toward a method or some research into this kind of problem.
2
-
$\begingroup$ cnn is all about supervised learning...define a score for each image and train accordingly $\endgroup$– user9947Jun 29, 2018 at 2:10
-
$\begingroup$ It would be difficult to determine how shiny a surface is in absolute terms without a reference value of some sort, but you could probably estimate the relative reflectance by comparing the apparent luminance of surfaces with the background luminance (I would expect shinier surfaces to appear brighter). However, this depends on a lot of factors, mainly orientation (if a surface is facing away from a light source, it would appear darker) and internal illumination (what is the reflectance of a light bulb when it is turned on?). $\endgroup$– cantordustJun 29, 2018 at 3:31
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
You should first know about the layer-wise working of a convolutional neural network. Read this https://distill.pub/2018/building-blocks/
Each layer of a cnn forms representations that are increasingly complex , which are combinations of simpler representations formed by its previous layers. "shine" is not a pattern (from image processing perspective). shine occurs due to change in intensity values.
And for quantifying "shine" you dont have to go for convolutional networks only , there are many more simpler methods for such tasks you could try fitting simple linear model and progress towards more complex ones.
and for classifying textures , there is interesting work going with "invariant scattering convolutional networks" you should check that.
answered Jun 29, 2018 at 7:41