Research on Natural Scene Vehicle Nameplate Text Detection Based on Improved DBNet

Abstract

Vehicle nameplate information as the main content of vehicle test, it is an important guarantee for the test quality of automobile testing institutions, and an important basis for the transportation authorities to determine the consistency of vehicle parameter configuration. Aiming at the problems of diverse text distribution, variable scale and complex background in vehicle nameplate detection, this paper proposes a dense connection and feature enhancement based on differentiable Binarization (DBNet) semantic segmentation algorithm. This algorithm uses the Dense Atrous Spatial Pyramid Pooling (DASPP) module to establish the connection between multiple dilated convolutions, capture dense sampling point pixels, and improve the utilization of high-level feature information. Secondly, the Feature Pyramid Enhancement Module (FPEM) is used to enhance the expression ability of the multi-layer feature information output from the backbone network, and the Feature Fusion Module (FFM) is used to fuse the feature information of different scales output from the FPEM, which improves the complementary ability between the features of each layer and obtains more comprehensive feature map information. Finally, the output of the DASPP and the FFM are concatenated to get the final segmentation results. The experimental results show that the improved algorithm can effectively locate the nameplate text area in the complex background. The detection accuracy on the self-defined datasets reaches 90.4 %, which is 2.6 % higher than the original algorithm DBNet.