A framework for synthetic dataset generation in pixel-wise crack segmentation

Description

Automated crack detection and segmentation for concrete and masonry surfaces has been a field of interest for many years now, as it can provide many benefits like early detection of structural building damage without the need for manual inspection. This task has traditionally been tackled using image processing techniques, but over the last few years neural networks have shown to be more effective. The downside to this approach however is that neural networks typically rely on large volumes of high-quality and diverse data for model training. This data is often expensive, time-consuming, and inconsistent to collect due to the manual annotation required. Especially for tasks like crack segmentation on masonry surfaces, which are more complex than the standard concrete segmentation, this becomes a problem.

In the past, the data collection issue has been solved by using 2D image generation approaches. One might be able to boost the performance of a neural network by padding an existing data set of real images with synthetic images such that the data set becomes bigger and more diverse. The key downside to these approaches however is their applicability: it is very hard to generate complex and realistic 2D crack images from scratch given a set of parameters such that they fit the function of the existing data without reusing the existing data. To this end, we propose a novel data set generation framework that can generate semi-realistic cracks in masonry surfaces in 3D. By generating a data set in 3D, we are able to use modern 3D modeling software to recreate realistic scenarios that can then be rendered into a complex and diverse data set. We show that through the use of partially synthetic data sets, we are able to match the segmentation performance of non-synthetic data sets for a segmentation network developed at the Hanze University of Applied Sciences in a previous study. By analyzing the effect of the different synthetic data sets on the model, insight is gained into how the data set affects the results of the network and how this method can improve performance in future work.

Period	7 Nov 2024
Event title	1st Interdisciplinary Workshop on Computer Vision Technologies for the Built Environment
Event type	Workshop
Location	Groningen, NetherlandsShow on map