CoreDeep: Improving Crack Detection Algorithms Using Width Stochasticity

TL;DR

CoreDeep tackles crack detection in images where challenging backgrounds blur crack boundaries and width variability across datasets complicates segmentation. It introduces a stochastic width (SW) augmentation that dilates ground-truth masks with $3\\times3$, $5\\times5$, and $8\\times8$ to model width variability, paired with RandAugment-based random augmentations and a binary focal dice loss to boost detectability. A thresholding step on pixel probabilities further refines predictions and reduces connectivity errors, achieving best results around $0.95$ on validation and improving mean IoU (mIoU) across ResNet50 and EfficientNet backbones on a merged crack segmentation dataset of about $11{,}298$ images. The approach yields substantial reductions in image-level false positives and false negatives, improves perceptual mask quality, and offers a practical, scalable solution for automated crack monitoring in infrastructure.

Abstract

Automatically detecting or segmenting cracks in images can help in reducing the cost of maintenance or operations. Detecting, measuring and quantifying cracks for distress analysis in challenging background scenarios is a difficult task as there is no clear boundary that separates cracks from the background. Developed algorithms should handle the inherent challenges associated with data. Some of the perceptually noted challenges are color, intensity, depth, blur, motion-blur, orientation, different region of interest (ROI) for the defect, scale, illumination, complex and challenging background, etc. These variations occur across (crack inter class) and within images (crack intra-class variabilities). Overall, there is significant background (inter) and foreground (intra-class) variability. In this work, we have attempted to reduce the effect of these variations in challenging background scenarios. We have proposed a stochastic width (SW) approach to reduce the effect of these variations. Our proposed approach improves detectability and significantly reduces false positives and negatives. We have measured the performance of our algorithm objectively in terms of mean IoU, false positives and negatives and subjectively in terms of perceptual quality.

Figures (6)

• Figure 1: The flowchart shows the process stochastic width augmentation. The input ground truth mask is augmented multiple times using Dilation 3, Dilation 5 and Dilation 8 incrementally. The augmented dataset will include 4 masks (Input Ground Truth Mask, Augmented Mask 1, Augmented Mask 2 and Augmented Mask 3) for each input.
• Figure 2: Shows (a) original image, (b) ground truth, (c) dilated with mask 3$\times$3, (d) dilated (c) mask 5$\times$5 and (e) dilated (d) mask 8$\times$8.
• Figure 3: The flowchart shows the process of calculating the threshold probability for each validation image as discussed in section \ref{['threshold selection']}. $idx$ is the bin index ($0$ to $9$) and $level$ is the number of times (maximum of 2 times) binning is done. $N\_current$ is the number of connected components in the original prediction and $N\_new$ is the number of connected components after applying a threshold on the original prediction. The calculated thresholds are analyzed to get the final threshold to use on test data (see section \ref{['threshold selection']}).
• Figure 4: All approaches here are trained using resnet50 backbone and test split of each dataset was used for predictions.
• Figure 5: All approaches here are trained using efficientnetb4 backbone and test split of each dataset was used for predictions.