BENCHMARKING PROBABILISTIC DEEP LEARNING METHODS FOR LICENSE PLATE RECOGNITION

Abstract

Automated license plate recognition (ALPR) systems assume training-test data alignment but struggle
under extreme conditions or in forensic scenarios without device-specific training. This study introduces explicit
modeling of prediction uncertainty to identify unreliable results. Three uncertainty quantification methods are
compared across two architectures. Experiments using synthetic noisy or blurred images demonstrate the
effectiveness of uncertainty in detecting errors. Additionally, a multi-task approach combining classification and
super-resolution improves recognition performance by 109% and error detection by 29%. These findings highlight
the role of uncertainty quantification in enhancing ALPR reliability, mitigating false identifications, and bolstering
system robustness in challenging operational environments algorithms can add more value to the customer retention
strategies.[1]