Automatic License Plate Recognition (ALPR) systems in real-world CCTV environments suffer severe performance degradation due to low resolution and complex non-linear degradations caused by long-distance capturing and varied weather conditions. Existing superresolution techniques are limited by focusing on pixel-level restoration, compromising structural character information, or failing to flexibly adapt to real environments due to fixed scaling factors and idealistic degradation assumptions. To address these issues, this paper proposes SAF-LPR, a novel stroke-aware invertible neural network framework. We introduce an "Invertible Weight Transfer" strategy to effectively model the physical inverse operation of the actual degradation process. The proposed two-stage approach first learns actual degradation patterns and noise via a deep residual pyramid encoder and utilizes the transposed filters as initial values for the restoration decoder. Subsequently, arbitrary scale rescaling and adaptive degradation modulation technologies are integrated, and finally, images optimized for recognizers are generated through a semantic feedback loop based on stroke attention and text priors. Experimental results on the real-world UFPR-SR-Plates dataset demonstrate that the proposed model significantly outperforms existing state-of-the-art models in both quantitative image quality metrics (PSNR) and recognition accuracy, proving its superior capability to restore clear structures even in severely damaged characters.
Lee et al. (2026) studied this question.