Methods for automated classification of pavement distress images are examined and compared. Images are divided into regions, and a two-stage classification procedure takes place. First, the regions are classified into primitives (plain, longitudinal, transverse, diagonal, or joint), which are the building blocks characterizing the various distress classes. The regional classification results are aggregated and used as input for the classification of the entire image to one of the classes of interest (plain, longitudinal, transverse, block, or alligator). A large number of features are examined using discriminant analysis, k-Nearest Neighbor, and discrete choice models. Conclusions are drawn on the discriminatory power of the various features and the appropriateness of the classification methods. For both levels, the logit model provided the best and most robust classification results. The sensitivity of the overall classification accuracy to the accuracy of primitive classification is also investigated.
Available at: http://works.bepress.com/allen_downey/3/