Since we primarily aim to recognize objects in images, why don’t we just use data from one of the image’s layers, or a grayscale image, as input and training data for the neural network?
I believe that just as the human can recognize objects in grayscale images, a neural network should be able to do the same. If we use all three RGB layers, I think the data from these layers are quite similar, which might lead to overfitting. Is that right?
No, it’s not correct.
You can use a grayscale image, this is true. But there is additional information available in a color image.
More information is usually a good thing.
There are other methods to control overfitting.
Hey there @zonehuang
Using grayscale images for object recognition is feasible and can reduce computational complexity, but RGB images provide more information about each channel (Red, Green, Blue) which can improve accuracy.
While grayscale might suffice for some tasks, RGB inputs often enhance performance and generalization by capturing a better set of features, thereby reducing the risk of overfitting compared to using grayscale.
Right! And of course it also critically matters what your goal is. If the stated goal is to distinguish between grey cats and brown cats, then greyscale images will not suffice. 
@zonehuang personally (and I am not sure-- have not tried/tested it) I do think you have an interesting idea here-- Insofar as the way ‘we’ do it, color adds some, but I would not say (necessarily) the ‘crucial’ component of what exists in the image, aside from certain particular classifiers as others have depicted.
So I wonder if you could do your train set on grayscale, which would probably save you a lot of time, and then later ‘add back in color’ via transfer learning.
I mean, upfront, you’d have the challenge your layers are not the same size (or perhaps you could just add null layers during the training ?).
No idea if it would work… Just speculating…