Crack Assessment Using Cascade Mask R-CNN and Dilation–Erosion Processing Technique

Abstract

Numerous studies have combined deep learning with image-based inspection to identify structural cracks. Some studies included the areas around the cracks in their labeling, requiring preprocessing for accurate quantification. Using this approach makes it difficult to generalize the thresholds used to distinguish cracked from noncracked regions. We propose a novel crack assessment technique that integrates the dilation–erosion method into deep learning frameworks for crack segmentation and introduce an evaluation method that leverages these outcomes. This method involves two phases: crack segmentation using a deep learning model and quantification via image processing. Training data were carefully labeled to include only crack interiors, and dilation was applied to expand pixel information. This improved training of the Cascade Mask Region-based Convolutional Neural Network (R-CNN) model, increasing the Intersection over Union by 7.53% across 20 pavement images. Applying erosion on the detection results yielded an average error of 0.018 mm in crack width, highlighting the method’s accuracy and precision.