Super-resolution for reliable industrial inspection under mixed degradations

Industrial defect inspection is sensitive to image degradation because blur, noise, compression, and limited spatial resolution can obscure subtle structural variations, while super-resolution (SR) reconstruction may introduce high-frequency artifacts that interfere with downstream inspection. To address this problem, we propose SynSR, an inspection-oriented SR framework for mixed degradations. Given a degraded low-resolution image, SynSR first performs structural prior extraction, where gradient and Laplacian responses are encoded into a compact prior to characterize local structural variations. The prior is then introduced into the restoration backbone through the Prior-Guided Feature Transfer Layer (PG-FTL), where a Global PriorGate adaptively modulates intermediate feature updates according to the structural prior, the current feature response, and optional degradation information. Before reconstruction, Artifact-Oriented Convolution (AOR-Conv) regulates intermediate responses in a spatial- and channel-aware manner to suppress unstable high-frequency activations that may propagate into the SR output. Different SR front-ends are finally evaluated under a unified SR-to-inspection protocol with matched test degradation conditions and unchanged downstream inspection architectures. Experiments on MVTec AD and MPDD show that SynSR achieves a favorable balance between reconstruction quality and downstream anomaly detection and localization performance under mixed degradations.