基于深度学习的实时车牌检测与识别

Real‑time license plate detection and recognition based on deep learning

深度学习已被广泛应用于车牌识别领域,但受限于光照条件对识别精度的影响,在夜间场景下的识别精度不高,且需要外界灯光的辅助。针对该问题,通过在数据预处理阶段引入随机暗化处理的手段,在原数据集的基础上模拟夜间效果,扩大样本数量,使得最终模型具有更好的光照鲁棒性。此外,因网络层数的不断增加,现有算法的识别速度很难满足实时识别,且对硬件性能提出了较高的要求。为在保证精度的同时有效提升算法的识别速度,通过构建二级级联定位网络Tiny-MTCNN(Tiny-Multi task convolutional neural network),结合关键点的初步区域提取和精确回归定位,减少网络层数。最终在测试环节,算法的平均识别精度达到96.5%,而在夜间场景下的精度显著提升至96.1%(提高2.7%)。另一方面,由于网络结构的优化,在精度略有损失(平均下降0.6%)的代价下,算法的识别速度显著提升了44.67%。

Deep learning has been widely used in the field of license plate recognition.However,recognitions in night scenes always depend on assistance of external lights.In order to solve this problem,this paper introduces a data preprocessing method named random darkening.This preprocessing method improves the size of the set by simulating night scenes on the original data set,and it brings the model a better lighting robustness.In addition,due to the increasing number of layers,the speed of existing algorithms is difficult to meet the requirements of real-time recognition.Therefore,a higher requirement for hardware performance is put forward.In order to improve the speed of the algorithm while ensuring its accuracy,this paper constructs a two-level cascaded positioning network Tiny-MTCNN(Tiny-Multi Task Convolutional Neural Network),which combines preliminary region extraction and precises regression positioning of key points.During the test,the average recognition accuracy of the algorithm reaches 96.5%,and the accuracy in the night scene is significantly improved to 96.1%(an increase of 2.7%).On the other hand,due to the optimization of network structure,the algorithm's recognition speed is significantly increased by 44.67%at the cost of a slight loss of accuracy(an average decrease of 0.6%).