Research On Traffic Sign Text Detection Method In Natural Scenes

The detection and recognition of traffic signs is one of the important research directions in the field of intelligent transportation systems,which has attracted the attention of many researchers.However,at present,research has mainly focused on the detection and recognition of various types of symbol-based traffic signs.There are few studies on text-based traffic signs,and the detection accuracy and real-time performance of related studies cannot meet the actual needs.In actual traffic scenes,there are a large number of text-based traffic signs,which provide rich and important road information in the form of text content.The efficient and accurate extraction of such text information plays an important role in intelligent assisted driving systems and unmanned driving technologies.This thesis combines the latest technology of image processing and deep learning,which is of great significance to carry out research on the traffic sign text detection in natural scenes.The main work and results of the thesis are as follows:(1)The image enhancement method based on Contrast Limited Adaptive Histogram Equalization is optimized and a data set for text-based traffic sign detection tasks constructed.Collected text-based traffic sign data sources in natural scenes,combined with affine transformation to expand the sample of the data set,and unified division and annotation of all images.In the process of constructing the data set,in order to reduce the impact of the natural environment on the detection effect,the character-based traffic sign image enhancement algorithm based on CLAHE is optimized.This algorithm improves the local histogram cropping threshold and color space selection scheme to make it stand out.At the same time,it retains the text details better and provides high-quality sample data for the study of character traffic signs.(2)A text-based traffic sign detection algorithm based on deep learning image semantic segmentation technology is proposed,which can accurately locate text-based traffic sign targets in natural scenes.Firstly,considering the detection speed at this stage,the traffic sign detection model uses a parallel network structure that combines lightweight network branches and shallow network branches;Secondly,by combining depth-wise separable convolution and dilated convolution,a lightweight space pyramid pooling module is proposed,which uses multiple dilated convolution branches to extract image features of different receptive fields to improve the perception ability of text-based traffic sign targets;Then the feature fusion methods of different branches are studied,and a feature fusion module based on attention mechanism is used to enable the feature maps of different branches to be adaptively fused,thereby improving the feature extraction ability of the network.The experimental results show that the textbased traffic sign detection algorithm proposed in this thesis can achieve a higher detection rate in different scenes and can obtain the target area more completely.(3)A text detection algorithm based on YOLO v3 is designed,which can realize rapid text detection in text-based traffic sign areas.Firstly,for the target area detected by the text-based traffic sign detection algorithm,a combination method based on LSDperspective conversion is adopted to achieve fast and effective area correction;Secondly,the size distribution of text targets in the text-based traffic sign area is analyzed,and it is found that the size of target recommendation box used in YOLO v3 is not suitable for the traffic sign text detection,so the K-means ++ clustering method is used to optimize the design of the target recommendation box.After that,the improved YOLO v3 algorithm for text detection is obtained.The experimental results show that the improved YOLO v3-based text detection algorithm mentioned in this thesis can effectively balance the accuracy and real-time performance of traffic sign text detection.