Abstract
Research in image super-resolution (SR), which seeks to enhance image quality by producing higher-resolution versions from low-quality inputs, has primarily focused on developing reconstruction algorithms rather than enhancing interpretability. SR networks continue to exhibit the opaque, black-box characteristics typical of deep learning, with limited research dedicated to investigating their internal mechanisms. This study aims to deepen the understanding of SR and conduct attribution analysis of SR networks from a holistic reconstruction perspective. We introduce a novel attribution method based on gradient and optical flow, termed GOFlow. Following verification with five SR models, we demonstrate that (1) GOFlow proves to be an effective tool for analyzing attributed pixels in SR neural networks from a comprehensive perspective; (2) Compared to the Layer Attribution Method (LAM), GOFlow produces a more detailed attribution map from a global perspective; (3) GOFlow is capable of exploring how texture and colorfulness influence the outputs of SR, proving that GOFlow can interpret SR models at the feature level.