To address poor feature extraction adaptability and the gap with practical engineering requirements for mountain bridge structural defect early warning under complex geological backgrounds and variable illumination, this paper proposes a CBAM-ResNet-Mask R-CNN instance-segmentation early warning framework for mountain-bridge disasters. A dual-path residual block is designed for the Stages 3-4 of the ResNet-101 backbone, where dilated convolution is adopted in the main path and 1×1 dimensionality-reduction convolution is introduced in the auxiliary path, so as to suppress complex background noise and enhance weak disaster features under uneven illumination. The CBAM attention module is embedded inside residual blocks of Stages 3-4 to realize joint channel-spatial feature screening. Since Mask R-CNN only outputs two-dimensional pixel-level segmentation masks from RGB images, the visual quantitative features calculated from segmentation masks are further fused with auxiliary structural monitoring data including vibration and strain time-series signals, to realize inversion of three-dimensional structural deformation and disaster severity assessment. The reduction rate of economic loss is estimated by feeding disaster-severity quantification results into a bridge life-cycle cost model. On the independent test set, the proposed model achieves mAP@0.5 of 92.17%, mAP@0.5:0.95 of 76.42% and multi-disaster comprehensive F1-score of 0.93. Its recall rate for small-scale rockfall early warning reaches 93.64%, with an early-warning false-alarm rate of 2.56%. The monitoring accuracy for bearing deformation and concrete spalling proportion both exceeds 90%, and the maximum economic-loss reduction rate reaches 58.46% under test-set conditions. The proposed framework can realize accurate disaster-feature segmentation and multi-source-data joint quantification for mountain-bridge hazards, which provides feasible technical support for mountain-bridge safety monitoring and graded early-warning decision-making.
