Abstract
By learning nonlinear mapping functions from low resolution (LR) images to high resolution (HR) images, deep neural networks show good performance in image super-resolution (SR). However, the existing SR approach has two potential limitations. First, learning the mapping function from LR to HR images is usually an ill-conditioned problem, since there exist an infinite number of HR images that can be down-sampled to the same LR image. Thus, the space of possible functions can be very large, making it difficult to find a good solution. Second, paired LR-HR data may not be available in real-world applications, and the underlying degradation method is often unknown. For this more general case, existing SR models tend to generate adaptive problems and produce poor performance. To solve the above problem, we propose a dual regression scheme that reduces the space of possible functions by introducing additional constraints on LR data.