Paper Info
Title
EGSANet: edge-guided sparse attention network for improving license plate detection in the wild
Abstract
The license plate (LP) has regular and prominent edge contour, which we believe is a key factor to improve the precision of LP detection. Therefore, we propose a novel Edge-Guided Sparse Attention (EGSA) mechanism, which is composed of an Edge-Guided Component (EGC) and a Sparse Attention Component (SAC) in tandem. EGC is designed as an end-to-end edge-preserving smoothing filter, which makes full use of convolutional neural networks (CNNs) and edge-guided map to enhance the edge contour of LPs and suppress noise effectively. On this basis, the K most relevant features adaptively searched by SAC are aggregated, thereby obtaining useful long-range dependencies. EGSA mechanism pays attention to the distinguishable regions and the important edge characteristics of LPs at the same time, so that the model can locate more precisely in the complex scenes. By evaluating state-of-the-art attention methods and LP detection models on both CCPD and AOLP datasets, we demonstrate the high precision and strong generalization capability of our proposed method.
Year
DOI
Venue
2022
10.1007/s10489-021-02628-4
APPLIED INTELLIGENCE
Keywords
DocType
Volume
License plate detection, Self-attention mechanism, Edge contour, Natural scene, Anchor-free
Journal
52
Issue
ISSN
Citations 
4
0924-669X
0
PageRank 
References 
Authors
0.34
0
4
Name
Order
Citations
PageRank
Jing J. Liang12073107.92
Guancheng Chen210.72
Yan Wang300.34
Huabiao Qin493.92