something wrong when we use encoder to process the high_frequency component

[zxp771]

Thanks for sharing your great work!

After decomposing the image (feature map) using the wavelet transform method you provided. We input the low-frequency and high-frequency components into an encoder network separately. After getting the encoded information, we found more detailed information in the high frequencies. Then, we have the following problem after re-inverting the encoded processed low and high-frequency components back to the time domain. A very distinct grid appears in the image (feature map). We have repeatedly experimented and found that coding operations on the low-frequency components do not cause this phenomenon. At the same time, this issue occurs only when coding operations are performed on high frequencies.

The processing of our work:

Could you please help us analyze the reason for this "grid effect"? Is it because the encoder changes the value of the high-frequency information so that the phase information in the high-frequency information changes? Does it end up causing the grid effect when inverted? If so, is there any good way to solve this problem?