rgb image input about recipes

[f0k]

You usually use three input channels (i.e., input_shape[1] == 3, as you suggested). It means the first convolutional layer will learn kernels that consider all three channels together (using different filters for red, green, blue that are added up afterwards), just like deeper layers in a ConvNet will consider all feature maps of the previous layer together.
See, for example, Andrej Karpathy's CS231n notes for an in-depth explanation.
You'll find several examples from literature in Lasagne/Recipes.