Search

Given a low-resolution, low-quality input image, super-resolution is the process of producing a high-resolution, high-quality output image. The input image is usually degraded by different factors, for example downscaling, random noise, or blurring resulting from the imaging sensor, lens, or motion. In addition, low-light enhancement is the process of producing an image with normal lighting, given an input image taken in poor lighting conditions. In order to combine these procedures in an end-to-end setup, one requires a dataset with low-resolution, low light images that are paired with high-resolution, normal-light images. The synthesized data that is usually generated, following a degradation process, is not necessarily identical to that of real-world scenarios. Recently, the availability of real-world low-light datasets such as RELLISUR, published at NeurIPS 2021, has made it possible to develop more realistic end-to-end low-light enhancement super-resolution algorithms based on real data. In this work, this dataset is used to train and evaluate an existing state-of-the-art model, namely EDSR. In addition, a comparison is performed to test the claim that end-to-end training yields better results than sequential processing for joint low-light enhancement and super-resolution. To handle low-light enhancement separately, the state-of-the-art ZeroDCE approach is chosen. Traditional methods are also used, such as bilinear and bicubic interpolation for upscaling, and histogram equalization for low-light enhancement.

Given a low-resolution, low-quality input image, super-resolution is the process of producing a high-resolution, high-quality output image. The input image is usually degraded by different factors, for example downscaling, random noise, or blurring resulting from the imaging sensor, lens, or motion. In addition, low-light enhancement is the process of producing an image with normal lighting, given an input image taken in poor lighting conditions. In order to combine these procedures in an end-to-end setup, one requires a dataset with low-resolution, low light images that are paired with high-resolution, normal-light images. The synthesized data that is usually generated, following a degradation process, is not necessarily identical to that of real-world scenarios. Recently, the availability of real-world low-light datasets such as RELLISUR, published at NeurIPS 2021, has made it possible to develop more realistic end-to-end low-light enhancement super-resolution algorithms based on real data. In this work, this dataset is used to train and evaluate an existing state-of-the-art model, namely EDSR. In addition, a comparison is performed to test the claim that end-to-end training yields better results than sequential processing for joint low-light enhancement and super-resolution. To handle low-light enhancement separately, the state-of-the-art ZeroDCE approach is chosen. Traditional methods are also used, such as bilinear and bicubic interpolation for upscaling, and histogram equalization for low-light enhancement.