Attention-Based, Multi-Scale CNN Algorithm for Automated Road Condition Assessment

Technology #11623

Key Points

  • Unique attention-based convolutional neural network (A+MCNN) provides automated procedures for classifying and segmenting various road surface objects
  • Can detect and classify 11 different kinds of road objects, including four distress classes (crack, crack seal, patch, pothole), five nondistress classes (joint, marker, manhole cover, curbing, shoulder), and two pavement classes (asphalt, concrete)


The University of Central Florida invention is a novel deep learning algorithm for automated (segmentation and classification of objects) road or pavement condition assessment. The A+MCNN has the following two unique features:

  • Scale‐specific feature extraction to produce feature maps of ROIs at three different scales
  • An attention module for mid‐fusion to produce score maps as weight matrices to determine how much each feature map at different scales contributes to final class label prediction