|
Abstract:In The technology of license plate detection and recognition is very important in modern transportation system. However, in the real scene, bad weather and oblique shooting angles will affect the accuracy of detection and recognition. Therefore, this paper mainly studies license plate detection and recognition algorithms under complex conditions.The license plate detection in this article uses the RetinaNet detection algorithm. First, build a ResNet network and FPN combination for image feature extraction and feature fusion, and then use Focal Loss regression to get an accurate prediction frame. The data is the CCPD data, which includs photos of various scenes, and the number is large, which can make the detection model more adaptable to complex conditions.The license plate recognition uses the combination of STN and LPRNet. STN is a kind of image space change network, which can perform affine change processing on the image. In this paper, the STN network is added to the LPRNet network, so that the network can learn how to reduce Loss through the spatial change of the image, and then the network can correct the slanted license plate. LPRNet is an end-to-end network that can recognize license plate characters without cutting characters. It is mainly composed of convolutional neural networks and CTC Loss. Finally, the effectiveness of the methodology of this paper is demonstrated through experiments. |