Research And Implementation Of Traffic Sign Text Detection And Recognition Based On Deep Learnin

With the development of artificial intelligence in recent years,many industries have seen the emergence of computer-aided production.In the field of transportation,autonomous driving has long been an important research project.In this paper,we study text-based traffic signs and use a deep learning approach to detect and recognize text information in text-based traffic signs.In text detection,the segmentation-based scene text detection algorithm DBNet is improved.The original DBNet is easy to lose some information in the feature extraction process,for this problem this paper incorporates the global attention mechanism in the downsampling stage of feature extraction,using the three-dimensional information of channel,spatial width and spatial height to better suppress non-text regions and improve the attention of effective regions,and replacing the convolution kernel of the Res Net network with a variable convolution.For the problem of large interval of text characters in traffic signs,the ASPP module is continued to be added on this basis to enhance the perceptual field of the network in the process of multi-scale fusion,so that the network captures more information to improve the accuracy of model detection.In text recognition,the CRNN text recognition network is improved.To make the model better extract Chinese features the original VGG network is replaced by the Dense Net backbone network,and a spatial transformation network is added during training to correct the detected text regions to improve the recognition accuracy.Considering the future deployment on mobile,a lightweight Mobile Net V3 network was used for comparison experiments.Finally,to solve the problem of insufficient dataset,this paper collects the national city names and some road information to make a traffic text dataset.Experiments show a 6.5% improvement in accuracy using the Dense Net network model compared to the original network,the use of Mobile Net V3 reduced the number of parameters by 67%.Finally,this paper constructs a platform for traffic sign text detection and recognition system and stores the recognition results in the database for other systems to call.