VorCrackSyn: An Efficient Voronoi-Based Method of Sequentially Synthesizing Crack Masks and Crack Images for Rapid Acquisition of Annotated Crack Data Sets

Deep learning (DL) enables intelligent and automated crack detection for civil infrastructure by learning from numerous pairs of crack images and labels. Pixel-level crack labels allow for quantitative crack severity assessments that can guide the maintenance priorities, thus playing an essential role in DL-based maintenance decision-making. However, acquiring pixel-level annotated crack data sets remains challenging due to the labor-intensive and time-consuming procedures of image acquisition and manual labeling. To overcome this challenge, this paper proposes VorCrackSyn, an efficient method that sequentially synthesizes crack masks and crack images. VorCrackSyn first synthesizes cracks with tortuous trajectories and varied widths by systematically integrating the developed methods of Voronoi-based partitioning, morphology-enhanced segmentation, and branching node-based path pruning. Subsequently, the synthetic cracks are integrated into real crack-free images via a background-guided random intensity method to produce realistic crack images for various infrastructures. A comparison method of crack severity rankings for ground truths and predictions is proposed to quantify the accuracy of the DL-based crack severity assessment. This method ranks the crack images by two severity metrics (equivalent crack length and width) and compares the ground truth and DL-based severity rankings. The results indicate that using the synthetic crack data sets of VorCrackSyn for pretraining can not only significantly enhance the DL-based crack detection performance but also drastically enhance the reliability of crack severity assessments under the condition of scarcity of real data. Additionally, VorCrackSyn can rapidly generate a 150-sample crack data set within 7 min on regular computers, effectively mitigating data scarcity.