Syncrack: Improving Pavement and Concrete Crack Detection through Synthetic Data Generation

Abstract

In crack detection, pixel-accurate predictions are necessary to measure the width -an important indicator of the severity of a crack. However, manual annotation of images to train supervised models is a hard and time-consuming task. Because of this, manual annotations tend to be inaccurate, particularly at pixel-accurate level. The learning bias introduced by this inaccuracy hinders pixel-accurate crack detection. In this paper we propose a novel tool aimed for synthetic image generation with accurate crack labels -Syncrack. This parametrizable tool also provides a method to introduce controlled noise to annotations, emulating human inaccuracy. By using this, first we do a robustness study of the impact of training with inaccurate labels. This study quantifies the detrimental effect of inaccurate annotations in the final prediction scores. Afterwards, we propose to use Syncrack to avoid this detrimental effect in a real-life context. For this, we show the advantages of using Syncrack generated images with accurate annotations for crack detection on real road images. Since supervised scores are biased by the inaccuracy of annotations, we propose a set of unsupervised metrics to evaluate the segmentation quality in terms of crack width. a https://orcid.

A controllable generative model for generating pavement crack images in complex scenes

“…In this section, the proposed CDM is compared with state‐of‐the‐art crack image generation methods, including the following: (1)Method 1: SynCrack (Rill‐García et al., 2022) uses the Perlin noise algorithm (Hart, 2001) to construct the background image, and adds the crack mask image to the image with a certain weight. (2)Method 2: CFC‐GAN (Sekar & Perumal, 2022) generates crack images with different brightness values based on the GAN framework. (3)Method 3: PG‐GAN (Chen et al., 2022) generates crack images with different shapes based on an improved GAN framework. (4)Method 4: VAE‐DCGAN (Pei et al., 2021b) combines the VAE method and the GAN framework, to generate different forms of crack images based on the hidden parameters of VAE. …”

“…(1) Method 1: SynCrack (Rill-García et al, 2022) uses the Perlin noise algorithm (Hart, 2001) to construct the background image, and adds the crack mask image to the image with a certain weight. (2) Method 2: CFC-GAN (Sekar & Perumal, 2022) In Figure 8, the result of Method 1 can be seen, which generates a background image based on Perlin noise that is quite different from the texture of real pavement captured by the camera.…”

“…While transfer learning has been applied to investigate deep residual networks for image‐based structural damage recognition (Gao & Mosalam, 2018), the models' abilities to generalize for unseen data may be constrained by the limited availability of training data. Therefore, generating a large number of crack images is crucial for expanding the data set and improving crack recognition and segmentation in complex scenes (Rill‐García et al., 2022; Sekar & Perumal, 2022). A common approach to augmenting the data set is to change the brightness, rotation, translation, and shear of images (Shorten & Khoshgoftaar, 2019).…”

Automatic UAV-based airport pavement inspection using mixed real and virtual scenarios

“…Because of this, the generation of synthetic datasets with automated ground truth generation has been proposed in multiple approaches in order to replace or complement real images. Rill-García et al [16] introduced the Syncrack tool, which allows users to generate synthetic annotated images of pavement and concrete textures, but the models trained with Syncrack data showed a considerably lower recall than those trained on real datasets. In [17], synthetic data augmentation was obtained by overlapping crack images from the CrackForest dataset [10] into the roads in the KITTI and Cityscapes dataset.…”

A Posteriori Deep Learning Segmentation Quality Estimation Based on Prediction Entropy

“…Most of them are based on the morphology of regions or evaluating how different are adjacent ones using, most of the time, their intensity. All these solutions are not of particular interest to this research because few of them are related to Deep Learning (DL) and the few of them based on entropy employ this concept with a particularly unusual meaning Pal and Bhandari, 1993;Zhang et al, 2004;Hao et al, 2009;Rill-García et al, 2022). They use entropy as a criterion to evaluate the intensity homogeneity inside each region (narrow histograms) before checking if the histograms of adjacent regions are disjoint.…”