A Real-Time and Efficient License Plate Detection and Alignment Algorithm Based on Neural Network

The detection and recognition of vehicle numberplate plays a vital part in intelligent traffic management. For numberplate detection in natural scenes, the accuracy of traditional methods is low, and the accuracy of license plate detection based on deep learning is high, while the real-time performance also is poor. For distorted license plate correction and alignment, a special network is used to return the four tops of the license plate, and then the transmission transformation [1] is carried out to correct the numberplate area, which further increases the computational complexity. For the existing model of detecting and returning the four vertices of the license plate, MTCNN [2] can be used as a reference. However, for images with larger input size, MTCNN has more image pyramids, the calculation amount is doubled, and the four vertices detected are not accurate. Therefore, a backbone network, called mobileSegnet for short, is proposed in this paper. The network learns from mobilenetV3 [3] and carries out up-sampling, and fuses the detailed features of the lower layer to segment the license plate area in the upper layer features. After the license plate area is segmented, the four vertices of the plate mask are found by traditional algorithm and the plate is aligned by transmission transformation. For different colors of license plates, different channels correspond to different colors of license plates in the output mask, which is conducive to the segmentation and recognition of double-layer license plates, and different colors of license plates can be expanded at will. Through a good deal of experiments, the detection of license plate using the segmentation task has higher precision and faster convergence than the detection and feature point regression task. And the slightly modified mobilenetV3 backbone network inference speed is faster, improving the reuse of the network layer, reducing the amount of subsequent numberplate correction and alignment calculation. Experimental results demonstrate that the proposed algorithm not only has high real-time performance and accuracy, but also has high use value in related fields.