# `Divergent Thinking`

Danger ！ Danger ！ Secret ！ Secret ！

Their downsampling operation just first divided the original image into non-overlapping patches with each size is 2 $\times$ 2. Is there a possibility, we could use a cascading network to denoise low light image, which divided the input image with patch size is 2 $\times$ 2 at first level and then processed the image for denoising. In next level, the network divided the input image with patch size is 4 $\times$ 4. And so on?
Is there a possibility, we could change the downsampling operation before using 2 $\times$ 2 convolution kernel process the downsampling patches?
Is there a possibility, we could use the two downsampling patches as PairLIE’s input and constructed P-Net, L-Net, R-Net and necessary regularization loss for them?
Why mapping the first downsampling patch into the second downsampling patch is equivalent to mapping the original noisy image into the clean image?

# Divergent Thinking