Benchmarking Probabilistic Deep Learning Methods for License Plate Recognition

Learning-based algorithms for automated license plate recognition implicitly assume that the training and test data are well aligned. However, this may not be case under extreme environmental conditions, or in forensic applications where system cannot trained a specific acquisition device. Predictions on such out-of-distribution images have an increased chance of failing. But failure is oftentimes hard to recognize human operator system. Hence, work we propose model prediction uncertainty explicitly. Such measure allows detect false predictions, indicating analyst when trust result recognition. In paper, compare three methods quantification two architectures. The experiments synthetic noisy blurred low-resolution show predictive reliably finds wrong predictions. We also multi-task combination classification super-resolution improves performance by 109% detection predictions 29%.