Enhancing YOLO deep networks for the detection of license plates in complex scenes

Abstract

License plate detection (LPD) in context is a challenging problem due to its sensitivity to environmental factors (such as rain, dust, and shadow) and light, which may greatly influence the detection accuracy. Moreover, LPD is more challenging for real-time systems. The usage of deep learning attracted the attention of researchers in recent years. It is being widely employed to solve classification and detection problems. Recently, for object detection, a new deep network was developed, namely, You Only Look Once (YOLO). We propose a YOLO-inspired adaptive solution with optimized parameters to enhance detection performance. To improve the detection process, the proposed solution employs "model generator" and "testing configurator" components where each model is trained using one single deep network. In addition to testing the newly designed solution using the UFPR-ALPR dataset, this work introduces a new annotated dataset for Canadian license plates (LP), namely CENPARMI datasets. The newly introduced dataset is challenging as it contains images with different settings in terms of: brightness, skewing, and distance. In addition to reporting the recall ratio results, a detailed error analysis to provide some insights into the types of false positives has been conducted. The proposed solution choice of optimal parameters enhanced the recall ratio and precision. For example, the proposed solution improved the recall ratio from 84.4% to 98.3% and the precision from 65.37% to 89.17% when tested using the UFPR-ALPR dataset.