Abstract:

Aiming at the problem that the existing bridge crack detection methods do not fully extract the rotation feature of bridge cracks and have low detection and segmentation accuracy， a bridge crack detection method based on improved Mask RCNN with rotation self-attention was proposed. Firstly， on the basis of the Mask R-CNN instance segmentation network， the ViTAE network based on Transformer learning is used as the backbone feature extraction network to improve the detection and segmentation accuracy of cracks. Then， a rotating variable window self-attention mechanism was designed to integrate into the bridge crack detection network to improve the detection ability of the feature extraction network for crack rotation features. Finally， the deformable convolution was used to further fit the irregular geometry of cracks to strengthen the recognition ability of crack feature information. Experimental results show that compared with the original Mask R-CNN detection and segmentation method， the accuracy of the proposed method is improved by 4.85%， the recall rate is increased by 13.95%， and the F₁-score can reach 91.66%. The proposed method can extract crack features more fully， achieve more accuratecrack detection， and is superior to the comparison methods in subjective and objective evaluation.

Key words: bridge engineering, concrete bridge crack, crack detection of diseases, rotational variable window attention, transformer learning