0
$\begingroup$

I was reading this well-known paper and noticed something in figure 1 below:

enter image description here

It says in the caption (The number of channels is denoted on top of the box). You can see that the number of channels is ranging from 1 up to 1024. I am confused here because it is known that the number of channels in colored images are 3 (R,G,B). Did I misunderstand something here?

Thank you.

$\endgroup$
1
  • $\begingroup$ This particular net has 1 channel as input (also called black and white image in the real world) $\endgroup$ Commented Jan 6, 2020 at 14:50

2 Answers 2

0
$\begingroup$

Here it's not with respect to the number of channels in an image, it's related to the number of out_channels you get after you have applied Conv operations.

  • in_channels is the number of channels in the input(generally for an image it's either 1 or 3 depending on the image data, for video it's 4 or more as well etc.)

  • out_channels Number of channels produced after the Conv operation on the given input.

$\endgroup$
2
  • $\begingroup$ I see. So it is the number of filters/kernels used on each layer. I never thought it is also called channels. It is a bit misleading. Thank you for clarification. $\endgroup$
    – Dave
    Commented Jan 6, 2020 at 12:22
  • $\begingroup$ Yep I agree it's indeed misleading. $\endgroup$
    – Aditya
    Commented Jan 6, 2020 at 13:15
1
$\begingroup$

Yes, Images can have more than 3 channels. Satellites routinely record multiple frequencies at once (for instance infrared). Normal monitors can't render that outright and you'll have to project those channels back to RGB. A simple way to do that is false colors.

False Color renders

These three false-color images demonstrate the application of remote sensing in precision agriculture: The left image shows vegetation density and the middle image presence of water (greens / blue for wet soil and red for dry soil). The right image shows where crops are under stress, as is particularly the case in fields 120 and 119 (indicated by red and yellow pixels). These fields were due to be irrigated the following day.

But convolutional nets take inputs (for instance RGB) and transform them to higher dimensional representations by applying several convolutions over the data. That is what you are referring to now. You can think of those bands as different filters that highlight features that we originally detected in the RGB-channels. They live inside the pipeline, and are then projected to a (typically lower dimensional) target (in this case 2)

So You can also input multi-band pictures, convolve those to even more channels.

$\endgroup$
1
  • $\begingroup$ Beautiful Answer (+1) $\endgroup$
    – Aditya
    Commented Jan 6, 2020 at 16:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.