Advanced License Plate Detector in Low-Quality Images with Smooth Regression Constraint

Improving the detection performance of license plate detectors in low-quality images is one of the core goals of license plate recognition community. While low-quality images encompasses various forms, this paper primarily focuses on two issues of blurring and distortion which have been empirically identified as the main challenges encountered by systems deployed in practical scenarios. And based on the designed Smooth Regression Constraint from the perspective of detection heads, we investigate and propose a novel License Plate Detector, namely SrcLPD. Specifically, we observe that corner-based detection head is robust to low-quality image with geometric distortion, while center-based regression methods are insensitive to blurred images. Therefore, we propose a novel soft-coupling loss to integrate the two types of bounding box regression strategies into a unified framework to instantiate smooth regression constraint. Extensive experiments verify that the proposed SrcLPD explicitly combines the advantages of the dual heads. SrcLPD outperforms 4 subsets of the CCPD benchmark, especially for blurry low-quality images, surpassing the previous method by 0.7% and achieving 97.6% detection precision.