Benchmarking Probabilistic Deep Learning Methods for License Plate Recognition

This study benchmarks probabilistic deep learning methods for license plate recognition (LPR), focusing on enhancing accuracy and reliability under real-world conditions. Utilizing a dataset of license plate images, the approach includes comprehensive preprocessing steps such as resizing, normalization, augmentation, and super-resolution to handle low-quality inputs. The dataset is split into training, validation, and testing subsets, with the test set emphasizing out-of-distribution (OOD) scenarios. The system employs convolutional neural networks (CNNs), probabilistic models like SR2 methods to estimate prediction uncertainty. A multi-task learning model is introduced to simultaneously perform LPR and image super-resolution, leveraging shared features for improved performance. Evaluation metrics include accuracy, precision, recall, F1-score, mean squared error, and novel uncertainty-based measures such as prediction uncertainty and error detection rate. Results demonstrate a 109% accuracy improvement with the multi-task model and a 29% increase in error detection via uncertainty quantification, highlighting the system’s robustness and practical value in uncertain environments.