|
Title |
A Study on Preprocessing Techniques to Improve High-Resolution Image Segmentation of Concrete Cracks Using U-Net++
|
|
Authors |
최소영(So Yeong Choi) ; 김일순(Il Sun Kim) ; 양은익(Eun Ik Yang) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.4.67 |
|
Keywords |
콘크리트 균열 분할; 데이터 증강; 미탐율; 패치 기반 전처리; U-Net++ Concrete crack segmentation; Data augmentation; False negative rate; Patch-based preprocessing; U-Net++ |
|
Abstract |
This study analyzed the effects of preprocessing techniques and data-mixing augmentation strategies on the semantic segmentation performance of high-resolution concrete crack images using a U-Net++ architecture with a ResNet34 encoder. As preprocessing methods, resizing and patch-based chipping were applied, and CutMix and MixUp were combined as augmentation techniques, yielding a primary set of six experimental scenarios together with two additional input-resolution variants, evaluated on the AI Hub concrete crack dataset with a resolution of 1920×1080 pixels. The results showed that, under an image-level evaluation at the original 1920×1080 resolution, chipping improved fine-crack detectability, demonstrating its advantage in suppressing missed detection. This result is attributed to the preservation of the original spatial resolution, which helped maintain the high-frequency morphological features of fine cracks. However, because the Dice includes background false positives introduced during patch reconstruction, high-resolution resizing exceeded chipping. Although CutMix and MixUp exhibited a regularization effect, they did not lead to a substantial improvement in the final segmentation performance. The dominant factor governing crack segmentation accuracy is the preservation of spatial resolution in the input data, and that preprocessing strategy design should be prioritized over the selection of data augmentation techniques.
|