A Generative adversarial learning strategy for enhanced lightweight crack delineation networks

Abstract

Traditional manual crack detection has been gradually replaced by unmanned aerial vehicles (UAVs) since automation and intelligence became the inevitable trends in routine bridge maintenance. Deep learning-based real-time crack detection is an important link in this automation process. However, due to the limitations of the field of view and airborne computer performance, it is challenging to balance crack detection accuracy and efficiency at the same time. To address this issue, a novel Generative Adversarial Network (GAN)-based strategy is proposed in this paper. Different from the traditional ways, the GAN-based strategy can introduce the morphological difference between the predictions and the manual labels into training process, further improving the network performance while ensuring detection efficiency. Three lightweight networks with different depths are designed based on the Dense Block to analyze the impact of the proposed method. Novel Fault-Tolerance (FT) indexes are proposed to reflect the morphological differences in predictions. Finally, the effectiveness and robustness of the proposed method are verified by the crack detection of highway bridge piers. Results show that the proposed method can effectively improve the detection scores of UAV-captured images under limited network parameters.