downsampling_in_deep_ISR

Deep image super resolution is the process of up-scaling images with deep neural networks. Downsampling is a critical step in this process, where a high-resolution image is converted to a low-resolution version in order to be processed by a deep learning model for training. The quality of the resulting super-resolution image depends heavily on the downsampling method used in creating the training set. In this research, we compare several different downsampling methods in the context of deep image super resolution. The explored downsampling methods include linear interpolation, bilinear interpolation, nearest neighbor interpolation, lanczos filter, bicubic interpolation, hamming window and box filter. We first provide a qualitative comparison of each downsampling method used on images with varying upscale factors, and then provide an empirical study with quantitative results to demonstrate that linear interpolation, the hamming window, and lanczos filter work best for training super image resolution models.

Download the PDF of our work