Asymmetric Loss / Coco Shot 8
We propose asymmetric loss functions that control the trade-of between over-blurring and leaving in residual noise in cases where the network cannot perform well. We show denoised images at multiple levels of the run-time slope parameter λ.
To put the results obtained with asymmetric loss functions into perspective, we provide naive versions of the image. This is an image constructed by blending the noisy input with the denoised image (without asymmetry), using to a global mixing parameter. The mixing parameter is optimized to minimize L2 distance between the naive image and the corresponding result obtained with asymmetric loss.
All images in this report are compressed due to file size constraints. This makes the subtle residual noise present in these images harder to appreciate. The authors' version of this material features images of higher quality.