自适应置信度阈值的非限制场景车牌检测算法

License plate detection algorithm in unrestricted scenes based on adaptive confidence threshold

针对车牌检测模型泛化性低,在智慧交通的不同应用场景中复用困难的问题,提出一种自适应置信度阈值的非限制场景车牌检测算法。首先,构建多预测头网络模型,利用分割预测头减少模型复用的预处理工作,利用自适应置信度阈值预测头提升模型的检测能力,并利用多尺度融合机制及边框回归预测头来提升模型的泛化能力;其次,采用可微分二值网络训练方法,利用可微分二值变换联合训练分类置信度及置信度阈值来学习模型参数;最后,利用连通感知非极大值抑制(CANMS)方法提升车牌检测的后处理速度,并引入轻量级网络ResNet18作为特征提取骨干网络,以减少模型参数量,进一步地提高检测速度。实验结果表明,在中国城市停车场数据集(CCPD)的6个不同限制条件特点的场景中,所提算法可获得平均99.5%的准确率与99.8%的召回率,并达到每秒70帧的高效检测速率,优于Faster R-CNN、SSD等锚框类算法的性能;在3个补充场景测试集上,所提算法对不同分辨率、不同拍摄距离、不同拍摄俯仰角等非限制场景下的车牌检测精度均高于90%。可见,所提算法在非限制场景下具备良好的检测性能及泛化能力,可以满足模型复用的要求。

Aiming at the problem of low generalization of the license plate detection model, which makes it difficult to reuse in different application scenes of smart transportation, a license plate detection algorithm in unrestricted scenes based on adaptive confidence threshold was proposed. Firstly, a multi-prediction head network model was constructed, in it, the segmentation prediction head was used to reduce the model reuse pre-processing work, the adaptive confidence threshold prediction head was used to improve the model detection ability, and the multi-scale fusion mechanism and bounding box regression prediction head were used to improve the model generalization ability. Secondly, a differentiable binary network training method was adopted to learn model parameters through differentiable binary transformation combined with the training of classification confidence and confidence threshold. Finally, the Connectivity Aware Non-Maximum Suppression(CANMS) method was used to improve the post-processing speed of license plate detection, and the lightweight network ResNet18 was introduced as the backbone network of feature extraction to reduce the model parameters and further improve the detection speed. Experimental results show that in 6 scenes with different constraints in Chinese City Parking Dataset(CCPD), the proposed algorithm can achieve the average precision of 99. 5% and the recall of 99. 8%, and achieves the efficient detection rate of 70 frames per second, which are better than the performance of anchor-based algorithms such as Faster Region-Conventional Neural Network(Faster R-CNN) and Single Shot MultiBox Detector(SSD). On the three supplementary scene test sets, the license plate detection accuracy of the proposed algorithm is higher than 90% in unrestricted scenes with different resolutions, different shooting distances, and different shooting angles of pitch. Therefore, the proposed algorithm has good detection performance and generalization ability in unrestricted scenes, and can meet the requirements of model reuse.