What is the best way to process images for network training?

Question

Neural networks

Good afternoon. What is the best way to process images for training? I saw options to note grayscale (which is understandable), but I’m more interested in why the dimension is increased, although the network takes a smaller size at the input (let’s say they increase to 150x150, and the network takes 28x28). What would you advise if the network is trained on sliced ​​characters, mostly letters and numbers.

Answer 1

There really is no point in resizing the image.
Except for those cases when you have a training sample - it has different sizes of pictures. Then you need to come up with some common average size and scale all the pictures before training. You can do this on the fly or offline.
With grayscale - there may be a catch. Different formulas for translating the RGB->Grayscale vector can give different results. Here we need to understand what we want to get. Just averaging over all channels. Or according to the formula of human vision where the green channel has the highest coefficient. In this case, the influence of green noise will be stronger on the result.