PROJECT TITLE :

Learned Image Downscaling for Upscaling Using Content Adaptive Resampler

ABSTRACT:

SR models based on deep convolutional neural networks have shown greater performance in recovering the underlying high-resolution (HR) images from low-resolution (LR) images acquired using predetermined downscaling algorithms.. Learned downscaling using content adaptive resampler (CAR) with care for the upscaling process is shown in this study. With the proposed resampler network, pixels on the downscaled picture are generated using content-adaptive image resampling kernels applied to HR input. A differentiable upscaling module (SR) is used to transform the LR result into its HR equivalent. The proposed framework delivers a new state-of-the-art SR performance by upscaling guided image resamplers that adaptively maintain precise information that is necessary to the upscaling by back-propagating the reconstruction error down to the original HR input across the entire framework. Results show that LR image quality is equivalent to classic interpolation based methods and that considerable SR performance gains are realised by deep SR models trained together with CAR models.