Traffic Sign Detection And Recognition Based On Data Augmentation And Dual-Model Structure

In recent years,with the rapid development of artificial intelligence,more and more industries are related with artificial intelligence.In these industries,Autonomous Vehicles technology is a popular research technology in recent years.In addition,it is crucial of the driverless field to detect the traffic sign,which naturally attracts much attention.First of all,the system that can detect the information about the traffic sign will transmit the traffic sign information to the unmanned vehicle.After that,according to the feedback information,the unmanned vehicle can adjust the decision of driving more quickly.In real scenes,due to the fact that the images collected by unmanned vehicles are often high-resolution perspective images,and the traffic signs themselves are small,the proportion of traffic sign targets appears smaller in the entire image.In the existing detection algorithms,the detection target proportion is relatively large,making it difficult to lose information in the feature extraction process.Therefore,when using traffic sign detection algorithm to detect high-resolution image,its accuracy and recall rate are low,and its robustness is weak.To fix these defects,this paper improves some traffic sign detection algorithms based on deep learning,and verifies the effectiveness of the improved algorithm through experiments.The main contributions are as follows:1、Aiming at the problem of the model that the resolution of the input image is too large,which causes huge computational resources,this paper proposes a target-centric data enhancement method.By obtaining the label information of the traffic sign in the dataset,it can locate the position of traffic sign in the image.Furthermore,it achieves the transfer from a high-resolution image to a low-resolution image by cropping the center of the image.In the meanwhile,a low-resolution image can be produced.This method can also effectively avoid the problem of losing some effective information owing to the using of image scaling.2、Aiming at the problem of its low accuracy and low recall rate when detecting high-resolution images in the traffic sign detection algorithm.Especially for small-sized traffic signs,the effect is even less obvious.Therefore,this paper proposes a dual-model network structure algorithm.In the process of testing,we first use the model one in the dual-model network structure to detect the original image,and get the rough position of traffic signs in the original image.Then,on the basis of the output of the model one,the model uses cropped new image as the input image for model two.Finally,the output of model two which can obtain the specific position and category of the traffic sign in the original image is used to restore the original image.3、In order to verify the effectiveness of the above method,the paper conducted a large number of comparative experiments on Tsinghua-Tencent 100 K and CCTSDB dataset of traffic signs.Through the analysis and comparison of experimental results,it has shown that the data enhancement method based on the object-centered method can effectively reduce the usage of GPU.At the same time,it also shows that the dual-model network structure algorithm can effectively improve the traffic sign accuracy and recall rate of high-resolution images,and slightly improve the detection of low-resolution images.