Multi-Resolution ResNet for Road and Bridge Crack Detection
Author(s)
Nayyeri, F
Zhou, J
Griffith University Author(s)
Year published
2021
Metadata
Show full item recordAbstract
In this work, we present a novel ResNet-like approach called MR-CrackNet to detect and localise infrastructural cracks of different sizes. In a series of consecutive cycles, low-level feature maps such as crack edges and boundaries are extracted and combined with multi-resolution high-level feature maps such as crack regions. Two separate streams are designed to carry these feature maps after each cycle: low-level stream carries crack feature maps in full resolution, and high-level stream send crack features through an encoder-decoder network to create feature maps in different low resolutions. At each cycle, two feature ...
View more >In this work, we present a novel ResNet-like approach called MR-CrackNet to detect and localise infrastructural cracks of different sizes. In a series of consecutive cycles, low-level feature maps such as crack edges and boundaries are extracted and combined with multi-resolution high-level feature maps such as crack regions. Two separate streams are designed to carry these feature maps after each cycle: low-level stream carries crack feature maps in full resolution, and high-level stream send crack features through an encoder-decoder network to create feature maps in different low resolutions. At each cycle, two feature maps are combined, processed and sent back through two streams for the next cycle. This model is trained and evaluated on a new crack dataset of 2, 532 images. Quantitative and qualitative results show that MR-CrackNet outperforms the baseline models with a clear margin and it is able to extract significant crack features and achieve highly accurate crack detection.
View less >
View more >In this work, we present a novel ResNet-like approach called MR-CrackNet to detect and localise infrastructural cracks of different sizes. In a series of consecutive cycles, low-level feature maps such as crack edges and boundaries are extracted and combined with multi-resolution high-level feature maps such as crack regions. Two separate streams are designed to carry these feature maps after each cycle: low-level stream carries crack feature maps in full resolution, and high-level stream send crack features through an encoder-decoder network to create feature maps in different low resolutions. At each cycle, two feature maps are combined, processed and sent back through two streams for the next cycle. This model is trained and evaluated on a new crack dataset of 2, 532 images. Quantitative and qualitative results show that MR-CrackNet outperforms the baseline models with a clear margin and it is able to extract significant crack features and achieve highly accurate crack detection.
View less >
Conference Title
DICTA 2021 - 2021 International Conference on Digital Image Computing: Techniques and Applications
Subject
Image processing
Engineering