Research on Real-time and Efficient Alignment Algorithm for License Plate Detection Based on Neural Network A lightweight network for license plate detection and alignment

Abstract

As one of the main identification marks of vehicles, vehicle license plate plays an important role in intelligent traffic management, and vehicle license plate detection and recognition is also a hot topic in recent years. Low accuracy of traditional method for license plate detection under the natural scene, the accuracy is higher based on the deep learning of license plate detection, but the real-time performance is poorer, for distortion correction of license plate alignment with special network return the four vertices of license plate, then the perspective transform <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]</sup> correct license plate area, the process to further increase the computational complexity. For existing detection and license plate four vertices of regression model, such as MTCNN <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[2]</sup> , but for the big input picture, MTCNN's image pyramid has more computation, so this paper proposes a mobilenet <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[3]</sup> backbone network and sampling at the three different scale layer on the license plate of the network to forecast the results, the network reference yolov4 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[4]</sup> design ideas, so called mobilenet - yolov4, in addition to regress the category of the license plate and boundary box. Four vertices regress mission add to the license plate detection. The integration of these three tasks into a simple network improves the multiplexing of the network layer and reduces the computation of subsequent license plate alignment. Experiments show that the algorithm has higher real-time and accuracy, and has higher research and application value.