Regularized Adaptation for Stable and Efficient Continuous-Level Learning on Image Processing Networks

In Convolutional Neural Network (CNN) based image processing, most of the\nstudies propose networks that are optimized for a single-level (or a\nsingle-objective); thus, they underperform on other levels and must be\nretrained for delivery of optimal performance. Using multiple models to cover\nmultiple levels involves very high computational costs. To solve these\nproblems, recent approaches train the networks on two different levels and\npropose their own interpolation methods to enable the arbitrary intermediate\nlevels. However, many of them fail to adapt hard tasks or interpolate smoothly,\nor the others still require large memory and computational cost. In this paper,\nwe propose a novel continuous-level learning framework using a Filter\nTransition Network (FTN) which is a non-linear module that easily adapt to new\nlevels, and is regularized to prevent undesirable side-effects. Additionally,\nfor stable learning of FTN, we newly propose a method to initialize non-linear\nCNNs with identity mappings. Furthermore, FTN is extremely lightweight module\nsince it is a data-independent module, which means it is not affected by the\nspatial resolution of the inputs. Extensive results for various image\nprocessing tasks indicate that the performance of FTN is stable in terms of\nadaptation and interpolation, and comparable to that of the other heavy\nframeworks.\n