Research on crack direction classification of bridge asphalt pavement based on machine vision

Bridge construction in China is gradually shifting from “construction-oriented” to “construction-supporting simultaneously”. Asphalt pavement is an indispensable part of bridge superstructure, and the appearance of cracks will shorten the service life of bridge pavement and affect the comfort and safety of driving. Therefore, how to quickly and accurately identify the cracks of bridge asphalt pavement is an important research topic. Therefore, based on the principle of traditional bilateral filtering and Retinex algorithm, a multi-scale Retinex algorithm with improved bilateral filtering is proposed in this paper to realize the de-noising enhancement of the crack image of asphalt pavement of bridge. Then the multi-threshold Otsu threshold method is used to improve the selection of K-means clustering center points, overcome the random instability of traditional K-means clustering center points, and improve the operation efficiency. Finally, MATLAB Deep Network Designer toolbox is used to train and compare 6 transfer learning models, such as GoogLeNet, ShuffleNet and MobileNet-v2. Considering the accuracy, loss function and training time, the GoogLeNet model is the best in the classification of the crack direction of asphalt pavement under the same hyperparameter conditions. The accuracy rate of the classification method used in this paper is 99.375%, which is conducive to the efficient and accurate identification of cracks in bridge asphalt pavement.