Adaptive Mask-Guided Crack Attention Mask R-CNN for Automatic Crack Detection

Crack detection plays an important role in infrastructure inspection; however, the utilization of existing pixel-level crack annotations for object detection remains challenging due to the absence of detection-oriented bounding-box labels. In this study, an Adaptive Crack-Guided Attention Mask R-CNN (Adaptive CGA-Mask R-CNN) framework is proposed for automatic crack detection. An adaptive mask-to-box generation strategy is developed to transform pixel-level crack masks into object detection annotations through background identification, morphological refinement, and connected-component analysis. Furthermore, an adaptive crack-guided attention module is introduced into the feature pyramid network of Mask R-CNN to enhance crack-related feature representation across multiple scales. The proposed framework preserves the original detection architecture while improving the sensitivity to thin, irregular, and complex crack structures. Experimental results demonstrate that the proposed method achieves competitive localization performance compared with conventional object detection approaches. Additional comparative and cross-dataset evaluations further verify the effectiveness and generalization capability of the proposed framework. The proposed method provides an efficient solution for exploiting existing segmentation annotations in automatic crack detection applications.