Automatic intelligent recognition of pavement distresses with limited dataset using generative adversarial networks

Automatic monitoring of pavement structure health has always been a significant problem for transportation engineers. Although the generative adversarial network (GAN) has proven to be an effective tool for improving pavement distress recognition accuracy, it may lead to increased computational cost, which inconsistent with the requirements of engineering practice. This paper describes a lightweight GAN structure for automatic pavement distress identification with high computation efficiency and low computation cost. Squeeze and expand (SE), multiscale convolution (MC), and depthwise separable convolution (DSC) were selected as alternative lightweight methods, and two series of comparative experiments were conducted. The results showed that the GAN-based model with SE implemented on its fully connected layer, MC&DSC implemented on its transpose convolution layers in the generator, and MC implemented on its convolution layers in the discriminator could reduce the largest proportion of model parameters (94.8%) while achieving satisfactory classification accuracy (85.4%).