Text‐Guided Crack Segmentation for Infrastructure Inspection: A Vision–Language Framework With Adaptive Multiscale Refinement

Automated crack segmentation is a prerequisite for scalable, objective condition assessment of civil infrastructure, yet conventional methods rely exclusively on visual features and remain vulnerable to low‐contrast surfaces, shadows, and background clutter encountered in routine inspections. This study introduces a text‐guided vision–language segmentation framework in which short, inspection‐style descriptions of crack type and surface condition, the kind of contextual knowledge inspectors already record, are used to condition pixel‐level prediction. Feature‐wise linear modulation (FiLM) aligns multiscale visual features with these linguistic descriptors, and an adaptive multiscale refinement module (AMRF) preserves crack boundary continuity under challenging surface conditions. To support this line of research, a multimodal dataset is contributed by annotating an existing crack mask repository with descriptive textual labels. On a unified benchmark combining seven public crack datasets, the proposed method achieves an F1 score of 79.37% and an intersection over union (IoU) of 65.80% when inspection descriptions are available, the highest among the convolutional, transformer, and foundation‐model baselines considered. It also shows a favorable precision–recall balance, faster training convergence, and qualitatively sharper feature activations in low‐contrast regions. All validation is limited to these seven public datasets; agency‐collected images and physical crack measurements were not evaluated. Beyond benchmark accuracy, the paper describes how the resulting masks can be converted into crack width, length, and density, which are inputs to established condition‐rating and maintenance‐prioritization workflows, and discusses deployment of the framework as an offline analysis module within existing inspection programs.