We have a talk in AIM2021
Prof. Chao Dong had a talk about Interpreting Super-Resolution Networks in AIM2021. The abstract of the talk is:
Although super resolution (SR) networks have achieved remarkable success in performance, their working mechanisms are still mysterious. Little attempts have been made in the interpretability of low-level vision tasks. In this talk, we will try to interpret SR network in three aspects – pixel, feature and filter. In the pixel level, we propose a new attribution approach called local attribution map (LAM). It could detect which input pixels contribute most for an output region. In the feature level, we have successfully discovered the “semantics” in SR networks, called deep degradation representations (DDR). We also reveal the differences in representation semantics between classification and SR networks. In the feature level, we develop a new diagnostic tool – Filter Attribution method based on Integral Gradient (FAIG) -- to find the most discriminative filters for degradation removal in blind SR networks. Our findings can not only help us better understand network behaviors, but also provide guidance on designing more efficient architectures.