Generalização de modelos de super-resolução em ALPR: um estudo sobre o impacto da degradação sintética no desempenho em domínios realistas

Abstract

Automatic License Plate Recognition (ALPR) systems face challenges when processing lowresolution images, in which the license plate area may represent less than 0.3% of the total capture, compromising optical character recognition (OCR). Super-Resolution (SR) emerges as a promising technique for reconstructing textual details prior to OCR; however, the literature lacks systematic studies on how synthetic degradation protocols — typically limited to bicubic interpolation with Gaussian blur — affect the generalization capacity of SR models to real-world capture conditions. This dissertation investigates two questions: (1) the impact of synthetic degradation protocol complexity on the cross-domain generalization of SR models; and (2) how the choice of perceptual loss function affects the semantic reconstruction of text on license plates. To this end, models based on three state-of-the-art architectures — RealESRGAN, ESRGAN, and LPR-RSR-EXT — were trained on five progressively complex degradation protocols, yielding 30 model configurations evaluated across 150 cross-domain scenarios, with performance measured by the character recognition rate using the YOLOv8 model. The results reveal a non-monotonic relationship between degradation complexity and generalization: models trained with two degradation stages recognized up to 18 times more characters in out-of-domain tests than models trained solely with Gaussian blur, whereas protocols with three or four stages reduced average accuracy by up to 59%, indicating destruction of semantic features due to excessive degradation. Regarding the perceptual loss function, architectural dependency was observed: ESRGAN with OCR Loss achieved 18.1% full plate match under the one-stage protocol, compared to 7.4% with VGG Loss, but collapsed completely (0.0% accuracy) with two or more degradation stages; RealESRGAN with OCR Loss reached an average of 1.22 recognized characters, against 3.56 with VGG Loss; and LPR-RSR-EXT, specialized in text, failed with VGG Loss (0.0%) but operated with OCR Loss (1.86 average characters). VGG Loss demonstrated consistent robustness across all evaluated conditions, establishing itself as a safe default choice regardless of the architecture or degradation protocol adopted.