A Boundary aware Graph attention framework for License Plate Detection under Challenging Traffic Conditions

Abstract

License Plate Detection (LPD) is one of the most important components of intelligent transportation systems, traffic surveillance systems, vehicular monitoring systems and automated law enforcement systems. LPD is complicated by various factors including lightening condition, occlusion, motion blur, complicated background condition and small scale license plates. In this research work introduced a new framework called Boundary-Aware Graph Attention Network (BA-GAT) to solve above mentioned problems in order to provide robust and accurate license plate detection under unconstrained traffic scenarios. The framework contains a Boundary-Aware Feature Enhancement Module in which Sobel and Scharr operators are utilized to enhance boundaries of license plates prior to deep feature extraction process. An improved CNN backbone using Channel & Spatial Attention mechanism (CBAM) is utilized to select discriminative channels and spatial parts and ignore the unnecessary background information. Moreover, Multi-Scale Feature Fusion Network is used to fuse hierarchical feature representations and detect small and long range license plates. To benefit from contextual relations between candidate regions, a Graph Attention Network (GAT) module is used to aggregate neighboring objects adaptively. Experimental results are reported on Kaggle License Plate Dataset, UFPR-ALPR, and CCPD datasets. The proposed BA-GAT model obtains precision of 97.8%, recall of 96.9%, F1-score.