Adversarial sample generation and training using geometric masks for accurate and resilient license plate character recognition
Makes license-plate recognition robust to deliberate tampering using geometric-mask adversarial training, lifting resilient accuracy from 25% to 99.7%.
Adversarially robust license-plate recognition
Abstract
Reading dirty or tampered license plates accurately in moving vehicles is challenging, and plates are often intentionally altered to avoid recognition. This work trains deep CNNs to 99.5% character-classification accuracy, then observes accuracy collapse to 25% under adversarial tampering. By enriching the dataset with geometrically masked images and retraining, the attack-aware model (AA-LPCR) recovers to 99.7% accuracy. Interpretability studies identify attack-prone regions of each character image, showing that embossed-plate systems can be upgraded to near-100% resilience.
Authors
Shrestha B, Khakurel G, Simkhada K, Adhikari B
arXiv preprint · cs.CV · 2023
How to cite
@misc{shrestha2023aalpcr,
title = {Adversarial sample generation and training using geometric masks for accurate and resilient license plate character recognition},
author = {Shrestha, Bishal and Khakurel, G. and Simkhada, K. and Adhikari, Badri},
year = {2023},
eprint = {2311.12857},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}