Automatic Pavement Crack Detection Based on Octave Convolution Neural Network with Hierarchical Feature Learning

Automatic pavement crack detection plays an important role in ensuring road safety. In
images of cracks, information about the cracks can be conveyed through high-frequency and low-frequency signals that focus on fine details and global structures, respectively. The output features obtained from different convolutional layers can be combined to represent information about both high-frequency and low-frequency signals. In this paper, we propose an encoder-decoder framework called octave hierarchical network (Octave-H), which is based on the U-Network (U-Net) architecture and utilizes an octave convolutional neural network and a hierarchical feature learning module for performing crack detection. The proposed octave convolution is capable of extracting multi-frequency feature maps, capturing both fine details and global cracks. We propose a hierarchical feature learning module that merges multi-frequency-scale feature maps with different levels (high and low) of octave convolutional layers. To verify the superiority of the proposed Octave-H, we
employed the CrackForest dataset (CFD) and AigleRN databases to evaluate this method. The
experimental results demonstrate that Octave-H outperforms other algorithms with satisfactory performance.