In CNN, How can I decide no. of feature map between input image and first hidden layer and so on?

Question

I have MxN rgb image. So here no. of color channels will be 3(r,g,b).now I want to construct a CNN model. so no of neurons input neurons will be MxN.Let we have nxn receptive fields.so no of neurons in first hidden layer will be (M-n+1)x(M-n+1).So my questions are:

1. How can I decide no. of feature map between input image and first hidden layer and so on?
2. How the feature map is connected with feature channel? is their any relations between feature channel and color channel?

I am new in CNN.any type of response accepted.

Answer 1

Answer 1: The number of feature map in any layer is a parameter you can tune. By having different value of these parameters (no. of feature maps) the model will end up learning different types of lower and higher level features. Best way to determine the number of feature maps is to try different values and check the accuracy (or visualize them).

Answer 2: In general, the previous layer feature maps (say, L1 number of feature maps) are considered as an L1-channel image and convolved with a valid kernel (also L1 channel kernel) and get a single feature map for the next layer. If you want to have L2 number of feature maps in the next layer, you basically have to have L2 number of different kernel (those L1-channel kernels). This is true for all convolutional layers.

Hope it helps!