AASPNet: Adaptive Attention-Based Symmetric Progressive Network for Weakly Supervised Image Super-Resolution

In recent years, deep convolutional neural networks (DCNNs) have made strong progress on single-image super-resolution (SISR). Most of them work in a fully supervised way: they learn a non-linear mapping from a low-resolution (LR) image to its high-resolution (HR) version. The main problem is that this needs ground-truth HR images for supervision, which makes these methods hard to use widely. To ease this problem, we propose a weakly supervised framework named AASPNet (Adaptive Attention-Based Symmetric Progressive Network). It has two parts: an upsampling process and a downsampling process. Each part is built from several training units, but only the first unit needs supervision. In other words, with only ×2 supervision, AASPNet can produce HR results at ×4 and ×8. The symmetry between the two parts is used in two ways. First, it gives consistency constraints that help recover more accurate HR images. Second, a cross self-supervision recovers HR images in a weakly supervised way, using both image and feature consistency between the upsampling and downsampling parts. We also add an adaptive attention strategy to pick out useful features, so the HR images get better. Experiments on several benchmark datasets show that our weakly supervised AASPNet reaches results that are close to state-of-the-art fully supervised DCNN methods.