Multilabel CNN Model for Asphalt Distress Classification

Abstract

One of the most challenging tasks in pavement management and rehabilitation is to detect and classify different distress types from images collected during field surveys. In this paper, a multilabel convolutional neural network (CNN) model for classifying asphalt distress is proposed. Unlike typical CNN models that classify a single object per image, the proposed model can detect and classify multiple distress types per image, without prior knowledge of the distress location. The model can classify the distress types into four categories: alligator cracking, block cracking, longitudinal/transverse cracking, and pothole. The proposed model was trained and tested on a real data set comprising 42,520 images using different pretrained architectures with various hyperparameter combinations. The results demonstrate the robustness of the proposed model and its potential for crack detection and localization using weakly supervised machine learning methods that can cope with partially labeled data sets.