How to divide low frequency and high frequency？

[ys-dpc]

First of all, feel the author's open source spirit, great idea!

I know that low and high frequencies are divided according to a given proportion, but how are they divided along the channel dimension? Are they divided randomly according to proportion? Or are there other pre-processing in it?

Thank you very much.

[novioleo]

@ys-dpc low frequency and high frequency is just "concept",in practical use,just avg pool,you can get the background which means low frequency data,and the origin image with specific small kernel convolution,you can get high frequency information of image.

[feixiangdekaka]

3-3 convolution represents acquire high frequency feature and 22 average pool means acquire low frequency feature, 22 is enough for low frequency? 44 or 66 pool size will be a better low frequency?