Research on Road Pothole Recognition Based on AlexNet Modeling and BP Neural Network

In this thesis, for the problem of road pothole recognition, based on image processing and deep learning ideas, a picture recognition model is established by determining the pixel value, RGB color pattern, image texture and other indexes with the objectives of accuracy, recall, precision and F1 score. Considering the limitations of traditional image recognition techniques in terms of classification accuracy and processing speed, this paper firstly carries out image feature extraction by AlexNet model and preprocesses the original 301 images. Not only the extraction of features, but also the visualization of feature extraction is realized. Considering the influence of different pixel versions and different recognition models on classification accuracy, the training set and test set are constructed according to different pixel versions, and then the BP neural network and XGBoost model are used for training. After the model training, the model performance is evaluated using the evaluation indexes such as accuracy, recall, precision and F1 score, and it is found that the model with 440x440 pixels outperforms that with 330x330 pixels, and the model with BP neural network outperforms that with XGBoost. After comparing the accuracy, the pixels of 440x440 are selected in the present study, and the training model using BP neural network is used for the classification prediction. The model achieves 99.0% accuracy on the test set.