Traffic Sign Text Detection And Recognition Based On Deep Learning

In recent years,the country is committed to applying deep learning technology to enhance the intelligence of traffic scenes to further promote the high-quality development of transportation market.On the one hand,traffic sign text,as a kind of natural scene text,faces many problems such as complex and variable background,uneven lighting and partial occlusion;on the other hand,traffic sign text brings great challenges to the detection and recognition process due to its own special characteristics,such as compact text arrangement and non-uniform orientation.As one of the indispensable tasks in applications such as autonomous driving,intelligent transportation,intelligent navigation and assisted driving,deep learning technology can meet the requirements of accurate and fast traffic sign text detection and recognition,therefore,using a method based on deep learning to detect and recognize the text in traffic signs,this research has important research value and significance for promoting the high-quality development of the transportation market.value and significance.This topic mainly studies the technical problems faced by the above traffic sign text detection and recognition,and designs an end-to-end traffic sign text detection and recognition model based on the analysis of multi-stage text detection and recognition technology,and conducts experimental validation of the related method on the self-built CTSD₃600 Chinese trafic dataset,and obtains a more satisfactory traffic sign text detection and recognition rate,which achieves the topic The objective is achieved.The main research works of the project arc as follows:（1）Construction of Chinese trafic sign text dataset.Since there are fow studies on traffic sign text in China,there is no dataset for Chinese traffic sign text.In this project,we screen and organize the traffic data sets publicly available in China,and carry out image acquisition in the actual environment to obtain the self-built traffic sign dataset CTSD₃600（China Traffic Sign Dataset₃600）.The dataset contains 3600 images,which will be used for the experimental validation analysis of traffic sign text detection and recognition in this paper.（2）Inaccuracies in traffic sign text detection are caused by factors such as tilting of the sign text due to the shooting angle during the acquisition process,and inconsistent alignment of text in the sign.To address the above problems,through experimental comparison,this topic improves the PAN++text detection recognition model.The text detection part uses a text clustering algorithm to correctly segment adjacent text;and a recognition feedback module is added to the network model to solve the problem that the detection module,recognition module and feature extraction module do not work well with each other.Experimental verification of the improved model shows that the accuracy of text detection in traffic signs is improved from 75.6%to 81.2%,which is a good improvement to the text detection part.（3）In the text recognition part,a text normalization processing step is added to make the text in both directions consistent.The detected text with different alignment directions is classified,and the vertical text is rotated and processed to ensure that the horizontal text is the same when the vertical text is processed in the feature sequence.The recognition head is a Seq2Seq model based on multi-headed attention.The encoder in the recognition head consists of a linear transform layer and a multi-headed attention layer,and the decoder consists of two LSTM layers and a multi-headed attention layer,which effectively fuses temporal features and visual features.Experimental validation on the dataset CTSD₃600 shows that the accuracy of the improved end-to-end traffic sign text recognition can be improved from 72.1%to 80.6%,which is a good improvement for the text recognition part.（4）In order to ensure the real-time of the overall network model,the network model of this topic adopts Resnet18 as the backbone network,and uses two FPEM modules to solve the problem of small sensory field of Resnet18 network.It improves the processing speed of the network for images on the premise of ensuring the accuracy of the model.After experimental verification,the processing time of the improved PAN++model for a single image in this project is 57.5ms.