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

Abstract

Industrial image super-resolution (SR) plays a crucial role in various industrial applications by generating high-resolution images that enhance image quality, clarity, and texture. The interpretability of industrial SR models is becoming increasingly important, enabling designers and quality inspectors to perform detailed image analysis and make more informed decisions. However, existing interpretability methods struggle to adapt to the complex degradation and diverse image patterns in industrial SR, making it challenging to provide reliable and accurate interpretations. To address this challenge, we propose a novel approach, Real Attribution Maps (RAM), designed for precise interpretation of industrial SR. RAM introduces two key components: the multi-path downsampling (MPD) function and the multi-progressive degradation (MPG) function. The MPD generates multiple attribution paths by applying a range of downsampling strategies, while the MPG incorporates random degradation kernels to better simulate real-world conditions, ensuring more accurate feature attribution. The final attribution map is derived by averaging the results from all paths. Extensive experiments conducted on industrial datasets, including IndSR, Wafer Maps, and Pelvis, validate the effectiveness of RAM. Our results show substantial improvements in several interpretation evaluation metrics and enhanced visual explanations that eliminate irrelevant interference. This work provides a powerful and versatile tool for explaining industrial SR models, offering significant advances in the interpretability of complex industrial images.