The optimal model of pothole road detection and recognition by machine vision

• 1、School of science,Liaoning Technical University,Liaoning fuxin,123000
• 2、School of mechanical engineering,Liaoning Technical University,Liaoning fuxin,123000

Abstract：Pothole road detection and identification is a computer vision task designed to identify roads with potholes through digital images. This is of great significance to the research and application of geological exploration, space science and natural disasters. This paper mainly studies the use of machine learning methods in data sets to identify and extract image sample data and generate training models. In order to distinguish the unfamiliar image from the normal road image or the pothole road image when the model is trained by this method.On this basis, transfer learning and time distillation are used to improve the recognition accuracy, recognition speed and classification accuracy. Three mathematical models, VGG-19, ResNet-18 and improved VGG-19, were established, and the parameters were continuously adjusted and optimized based on the training model, and the recognition rate, speed and classification accuracy of the model were continuously improved. According to the results of code operation, the training model is evaluated from the aspects of accuracy, robustness, efficiency evaluation, generalization ability, interpretability, comparison experiment, usability and so on.Pyhton software was used to analyze the time efficiency, accurate loss value and model size through VGG-19, ResNet-18 and improved VGG-19 model, and finally the most accurate prediction was obtained. The results are as follows: the training time of the improved VGG-19 model is 31176.55s, the size of the training model file is 100MB, the lost value can converge to about 2.1×10-4 after 200 training times, and the accuracy rate of the test model can reach 98%. In the experiment, the average accuracy rate is about 97% after several experiments.The analysis results show that the error rate of the shadow pothole area is lower, and the working efficiency is higher, the generalization ability is stronger, and the interpretation is stronger. The evaluation results of the improved VGG-19 model show that the model has the highest accuracy in the pothole road detection and recognition accuracy index.

Keywords： optimization VGG-19 ResNet-18 Improved VGG-19