Recognition Of Traffic Signs In Natural Scenes Based On Small Target Detection

Traffic sign detection is an inevitable difficulty in autonomous driving and assisted driving technology.With the rise of deep learning,many researchers use deep learning to extract more robust features,and use deep learning-based detection algorithms to solve the problem of traffic sign detection.However,when the size of the target is smaller than that of the whole image,most of the detection algorithms are not good.However,the scene of automatic driving is complex,and the traffic sign is a small target in many cases.Therefore,it is necessary to improve the small target scene of traffic sign detection.The work in this paper is based on the traffic sign data set Tsinghua-Tencent 100 K,TT100K for short.This data set was collected and produced jointly by Tsinghua University and Tencent.The dataset contains a large number of natural scenes with small target traffic signs.However,this data set also has some defects of imbalanced categories.Therefore,this paper enhances and expands the data set,using appropriate data collected from the network or collected by ourselves to expand the data,using traffic sign template to supplement the sample number of categories with a small number of samples,so as to achieve category equalization.This paper mainly uses YOLOv3 detection algorithm for traffic sign detection,and proposes two main ideas for small target detection.The first is to use super-resolution to perform super-resolution reconstruction on the image,and the second is to improve the scale of the features predicted by YOLOv3.Specifically,because the proportion of small targets in the picture is small,the picture is relatively blurred,which leads to a large loss of semantic information and affects the detection effect.Therefore,this paper proposes to perform super-resolution reconstruction of the input picture before detection,improve the resolution of the area where small targets exist,and increase the semantic information of small target traffic sign objects,thereby improving the detection accuracy.In addition,this article has improved the YOLOv3 algorithm itself,using features that are more low-level and larger in size than the original algorithm for more detailed detection.However,increasing the resolution of the image will have a negative impact on the detection time.For this reason,a target determination network is also designed in this paper.By combining the image segmentation strategy,a large number of areas without traffic signs are excluded,reducing the amount of calculation.At the same time,the hyperparameters of the method in this paper can be adjusted to balance the running speed and detection accuracy.The segmentation method and NMS strategy are improved for the target segmentation problem,and repeated detection of the target segmentation is successfully avoided.And missed issues.Based on the above method,this paper uses super-resolution and large-scale feature prediction ideas to carry out detailed experiments on the TT100 K data set.Experiments show that the super-resolution idea successfully increases the semantic information of small-sized traffic signs and improves the detection effect of small target detection.In the end,the detection method proposed for small target traffic signs in this paper has achieved high evaluation indexes on the experimental data set as a whole,and has achieved excellent detection results.