Automated Pixel-Level Pavement Crack Detection on 3D Asphalt Surfaces with a Recurrent Neural Network

For automated crack detection on an asphalt pavement at pixel-level, the authors propose a recurrent neural network called CrackNet-R. For recursive updating of the internal memory of CrackNet-R, the authors also propose a gated recurrent multilayer perceptron (GRMLP). Compared to the widely used long short-term memory (LSTM) and gated recurrent unit (GRU), GRMLP yields deeper abstractions on inputs and hidden states by performing multilayer nonlinear transforms at gating units. CrackNet-R carries out a two-phase sequence processing: 1) sequence generation, that finds the local paths most likely to form crack patterns; 2) sequence modeling, that predicts the probability of the input sequence being a crack pattern. In sequence modeling, GRMLP slightly improves on LSTM and GRU by using just one more nonlinear layer at each gate. In addition to sequence processing, the authors propose an output layer to produce pixel probabilities. The output layer is crucial for pixel accuracy, as it enables the transition from sequence-level to pixel-level learning. Using a database of 3,000 three-dimensional images, CrackNet-R is trained by the optimization of sequence modeling, sequence generation, and the output layer. An experiment using 500 test pavement images demonstrates that CrackNet-R can simultaneously achieve high precision (89%), recall (95%), and F-measure (92%). CrackNet-R is four times faster than the original CrackNet and produces demonstrable improvements in the accuracy of pixel-level crack detection.


  • English

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01783780
  • Record Type: Publication
  • Files: TRIS
  • Created Date: Sep 29 2021 9:34AM