Deep learning and image preprocessing-based crack repair trace and secondary crack classification detection method for concrete bridges

Abstract

As a main type of vulnerability, a large number of cracks exists in concrete bridges, and some cracks will be secondary dehisced after repair, and the crack repair traces (CRTs) or the secondary cracks (SCs) are easily confused with concrete spalling and other defects during intelligent disease detection. Therefore, accurate classification of CRT and SC is crucial for effective detection of concrete bridge diseases. To achieve precise classification, the images of CRT and SC will be processed. Initially, images undergo Poisson-noise and Bilateral-filtering. Subsequently, image segmentation is performed using the OTSU algorithm. The resulting binary images, along with the unprocessed original images, are utilised for training the YOLOv7 target detection network. Finally, the trained model is verified by a test set containing negative samples. The training results show that after several rounds of iterative training, the binarized image detection model achieves exceptional performance, accurately classifying and detecting CRTs or SCs with an accuracy rate of 98.4%. In contrast, the accuracy rate for the original image is only 39.9%. The inference results obtained from the test set further validate the method’s rationality and feasibility, with a significant reduction in misclassification probabilities observed after 200 rounds of iterative training.