Ops, we did it again! Image denoising just leveled up!!! arxiv.org/abs/2006.06072
Our new method can do so much more than just predicting a denoised image!
Let us take you on a short tour… (1/6) #CARE #denoising #uncertainty #diversity @Mangal_Prakash_ @sagzehn @jug_lab
Our new method can do so much more than just predicting a denoised image!
Let us take you on a short tour… (1/6) #CARE #denoising #uncertainty #diversity @Mangal_Prakash_ @sagzehn @jug_lab
What always bugged us with any CARE method is that a single restored output is created, giving the false impression that the input can unambiguously be restored. But some information gets lost in the noise, so this can obviously not strictly be true! (2/6) arxiv.org/abs/2006.06072
A #DivNoising network is trained to know how structures in the data look like and can be used to predict diverse denoised outputs. In areas of the image where the input is ambiguous, #DivNoising will sample the space of reasonable interpretations.(3/6) arxiv.org/abs/2006.06072
With #DivNoising we propose a fundamentally new way to train neural networks for image denoising, using VAEs. No additional training data needed, only noisy data is enough! The required noise model can also be bootstrapped from the noisy input data… (4/6) arxiv.org/abs/2006.06072
Check the preprint to see that this all makes sense also from a more formal point of view. We are really excited about this and are sure that #DivNoising is powerful! We decided to demo its usefulness on an instance segmentation task. (5/6) arxiv.org/abs/2006.06072
A single high quality prediction can still be created, and such #DivNoising results are often better than any other unsupervised method (bold), at times even beating CARE (underl.), which has access to clean (ground truth) images during training. (6/6) arxiv.org/abs/2006.06072
Fun fact: working on #DivNoising was an intense and rewarding team effort together with @Mangal_Prakash_ and @sagzehn from the @jug_lab. But a linear author order just doesn’t capture how we work, so we made Alex first and last author... arxiv.org/abs/2006.06072
Bonus: the networks we used in the manuscript are tiny, most of them fitting in 2GB of your GPU. Hence, if you give us some time your Fiji might start talking #DivNoising to you... 😊 @FijiSc arxiv.org/abs/2006.06072
