This research investigates whether leveraging multi-temporal data for crack segmenta- tion can enhance segmentation quality. Cracks represent one of the earliest indicators of deterioration in concrete structures. Early automatic detection of these cracks can signif- icantly extend the lifespan of critical infrastructure such as bridges, buildings, and tun- nels, while simultaneously reducing maintenance costs and facilitating efficient structural health monitoring. The research compares a Swin UNETR trained on multi-temporal data with a U-Net trained on mono-temporal data to assess the effect of temporal information compared to conventional single-image approaches. To this end, a multi-temporal dataset comprising 1356 images, each with 32 sequential crack propagation images, was created. After training the models, experiments were conducted to analyze the generalization abil- ity, temporal consistency, and segmentation quality of both models. The multi-temporal approach consistently outperformed its mono-temporal counterpart, achieving an IoU of 82.72% and an F1-score of 90.54%, representing a significant improvement over the mono- temporal model’s IoU of 76.69% and F1-score of 86.18%, despite requiring only half the trainable parameters. Both models were further tested on a data-augmented test set sim- ulating environmental conditions, with the multi-temporal model again exhibiting su- perior performance and generalization ability. The multi-temporal model also displayed more consistent segmentation quality, with reduced noise and fewer errors. These results suggest that temporal information significantly enhances the performance of segmenta- tion models, offering a promising solution for improved crack detection and long-term monitoring of concrete structures, even with limited sequential data.
The code used in this work is split up into three directories within this repository. Refer to the README files in the directories for descriptions of the code.
models: Contains the main code for the U-Net and the Swin UNETRdata creation: Contains the code necessary to create the mono- and multi-temporal datasetcode: Preliminary work on the Retina dataset for pretraining. The idea was discarded later and the code in this directory is not used for this work.presentation: This directory contains the slides created to track progress. Later the final presentation will be added here.figures: All figures are in the submission report and were not uploaded to GitHub.