RAM: Interpreting real-world image super-resolution in the industry environment

Ze-Yu Mi, Yu-Bin Yang

Abstract: Highlights•We propose real attribution maps (RAM), a novel interpretability method tailored for real-world super-resolution (SR) tasks in industrial environments.•We identify the limitations of existing industrial interpretability methods and propose two key functions to adapt them to complex degradations and diverse image patterns in real-world scenarios.•We demonstrate the RAM’s superior performance over existing interpretability methods, accurately identifying critical regions in industrial SR and classification applications.