.png){kind=icon}
.png){kind=icon}
Return
Crack segmentation using discrete cosine transform in shadow environments
.png){kind=icon}
PRE
AI
DOI：10.1016/j.autcon.2024.105646.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate pavement crack segmentation is crucial for quantifying the extent of pavement damage. However, shadows from roadside trees or buildings often significantly impact the results of crack segmentation in actual crack detection or segmentation processes. To solve this problem, this paper presents a pavement crack segmentation network called SCSNet based on discrete cosine transform. SCSNet combines a proposed shadow removal module with a loss function based on pixel frequency distribution to further minimize the impact of shadows on accuracy. Additionally, a crack dataset with shadows was introduced. By comparing with the classical semantic segmentation network, the results show that SCSNet with training from scratch, outperforms classic models based on pre-trained weights. The result of the ablation experiments also demonstrate that the proposed tricks are effective. Finally, the actual crack segmentation results further demonstrate the superiority of SCSNet in segmenting cracks in complex environments.
Keywords:
Non-destructive testing
Deep learning
Crack segmentation
Shadow removal
Journal
IF:
11.5
Papers:
6.2K
Citations:
4.2W