| There is rich semantic information in traffic signs,which plays an irreplaceable role in traffic-related fields.In recent years,a series of highly intelligent concepts and products related to transportation,such as smart transportation,intelligent navigation,and autonomous driving,have received widespread Attention.However,at present,most of the studies on traffic signs are focused on extracting the symbolic information in traffic signs,while there are relatively few studies on the textual information in traffic signs.Therefore,traffic sign text recognition has great research significance and application value.Considering the current shortage of relevant public data sets and the characteristics of domestic Traffic signs.We made a set of text type traffic sign detection data sets in the wild:China Traffic Sign Text Detection Dataset(CTSTDD),a set of traffic sign text detection in any direction:China Traffic Sign Text Detection Dataset(CTSTDD),and a set of chinese traffic sign text recognition dataset:China Traffic Sign Text Recognition Dataset(CTSTRD).These self-made datasets fully consider various domestic scenarios,including various road in highways,urban areas,and rural areas,and the complex background of traffic signs in the data set and the large span of target scale all bring high challenges to the research work.In this paper,we design a process from rough to fine for traffic sign text recognition,which includes text type traffic sign detection,text detection and text recognition.The specific research contents are as follows:(1)we need to perform traffic sign detection to locate the location of traffic signs in the picture and remove a lot of irrelevant areas.In this paper,a NSYOLOv3 network suitable for traffic sign detection tasks in the wild is designed and manufactured.This network uses YOLOv3 as a basic model,and slims down the model for traffic sign categories to reduce the amount of parameters and calculations.Based on the experimental analysis on the CTTSDD data set,NSYOLOv3 achieved93.28% accuracy on the CTTSDD test data set,which is basically the same as the accuracy of YOLOv3,but the model parameter amount is only 13.58% of YOLOv3.(2)We perform text detection on the traffic sign detected in the previous step.Based on the classic EAST algorithm,this article introduces deformable convolution modules and irregular convolutions such as 1 * 7,7 * 1,etc.,based on the characteristics of oblique deformation,occlusion,and maximum aspect ratio of traffic sign text in the wild.We design and implement the TSEAST algorithm suitable for traffic sign text detection in the wild.The comprehensive index F on the CTSTDD test data set is 91.92%.Compared with the classic EAST algorithm,the TSEAST algorithm improves the accuracy slightly.The recall rate has been significantly improved,and the model size of TSEAST is only 6.2MB,which is 39.24% of the size of the EAST model.(3)Text recognition is performed on the text lines obtained by the text detection.In this paper,two classic text recognition algorithms are compared to select algorithms suitable for traffic sign recognition in the wild. |
