Research On Traffic Sign Recognition Based On Improved YOLOv3

Aiming at the problems of low accuracy,slow speed and large amount of model parameters in the current traffic sign recognition algorithm,this paper improves the YOLOv3 target detection algorithm,builds a new target detection model,and tests it under the expanded TT100 K data set.The main work contents are as follows:Firstly,aiming at the problem of low accuracy of target detection algorithm,ciou loss is used to replace the mean square error loss in the original YOLOv3 network,which optimizes the boundary box positioning loss and improves the problem of boundary box scale sensitivity.The mesh activation function is used to replace the leaky relu function in the original network to optimize the effect of gradient descent.Increasing the attention mechanism hardly changes the recognition accuracy of the network model on the premise of reducing the training parameters.Increase multi-scale prediction to effectively improve the detection effect of small targets.Integrating the above improvement points,the YOLO-A detection model is trained under the expanded TT100 K data set.After testing,the recognition accuracy of this model is 96%,which is13% higher than that of YOLOv3 model.Secondly,in view of the problem of slow algorithm speed and large model,this experiment uses deep separable convolution to replace the conventional convolution in the original network,greatly reducing the training parameters of the network model,and the YOLO-B detection model is trained under the expanded TT100 K data set.Compared with YOLOv3 model,the new detection model achieves the recognition speed of 42 frames per second,22 frames per second higher than that of YOLOv3 model.Meanwhile,the model size is only 89.6MB,which reduces the number of parameters by about 62% compared with YOLOv3 model,and successfully achieves the lightweight of network model.Finally,the YOLOv3-Ours detection model is obtained by integrating all the improved points and training on the expanded TT100 K data set by integrating the recognition accuracy,recognition speed and model size.The recognition accuracy of the new detection model is 94%,3% higher than that of the YOLOv5 s model,maintaining the recognition speed of 38 frames per second.4 frames per second higher than the SSD model.The improved algorithm in this paper can complete the target detection task with high precision and real time,and also provides the possibility for deployment on mobile platform.