An enhanced crack segmentation method using implicit classification and inference rules in steel bridge

In order to further improve the accuracy of crack segmentation of steel box girder bridges in complex backgrounds, a crack detection method based on information fusion combining implicit feature classification and masked image inference is proposed. Firstly, the crack segmentation model is employed to obtain the preliminary crack detection results from object recognition perspective. Then, the crack segmentation effect is enhanced with classification and inference rules for cracks learned from the training image set. In the novel masked inference module, the morphological correlation of cracks between different patches is automatically learned to provide the crack distribution probability. Meanwhile, the classification rules for cracks and distractors at the patch scale are captured and applied to suppress disturbances in classification module. Through iterative detection, inference and fusion of crack information at pixel scale and patch scale, the integrated segmentation network obtains the final crack identification results. Results from actual bridge images show that the crack detect effect of the enhanced crack detection integrated algorithm is significantly better than many classic single segmentation networks such as Deeplabv3Plus and Transformer. The average IoU, precision and recall of the steel bridge images in complex backgrounds reached 77.66%, 85.76% and 89.33%, respectively. The interference of objects such as handwriting, calibration rulers and welds can be effectively reduced.