Research On Image Recognition Method In Traffic Field Based On Neural Network

Due to the rapid development of deep learning,deep learning models have gradually become the mainstream tools for research in the traffic field.The number of layers in deep learning models is increasing,as is the computational complexity and volume,resulting in poor real-time performance.Additionally,the poor quality of traffic sign images acquired in low-light environments results in low recognition accuracy.To address the preceding two issues,the recognition method of the Chinese traffic sign dataset is investigated using deep learning as the theoretical foundation.The following are the main research findings:To begin with,people want to improve the accuracy of traffic sign recognition,which causes the model layers to become too deep,increasing computational complexity and volume,resulting in poor real-time performance.The YOLOv5 traffic sign recognition method is proposed,which uses model pruning and network quantization techniques to compress the model.The results show that when compared to the YOLOv5 model,the volume of model compression YOLOv5 decreases by 87.4 %,the number of parameters decreases by 88.2%,and the inference speed of GPU and CPU increases by74% and 75%,respectively,improving the real-time performance of traffic sign recognition.Second,to address the issue of poor real-time performance,a lightweight YOLOv5 traffic sign recognition method is proposed,in which the Stem module and the base unit network of Shufflenetv2 are introduced to replace the YOLOv5 backbone network.The lightweight YOLOv5 model reduces the number of parameters by 95.4% and the actual memory space by 93.9% while maintaining recognition accuracy of 95.9%,and inference speed on GPU and CPU increases by 79.7% and 75%,respectively,improving real-time traffic sign recognition performance.Finally,the image quality obtained by the automatic driving environment sensing device is poor in low light environments such as wind,frost,rain,snow,or evening,making it difficult to recognize traffic signs.To address this issue,a low light traffic sign recognition method is proposed,in which low light traffic sign images are enhanced by the RetinexNet model,and the Regnet model is improved for light enhanced traffic sign image recognition.The experimental results show that the low-light traffic sign recognition method outperforms the traditional deep learning model with 99.67%accuracy.Figure 33;Table 8;Reference 43...