Task-Driven Super Resolution for Improved Crack Segmentation in Low-Resolution Imagery

Currently, crack images captured by unmanned aerial vehicles are extensively used in the field of crack detection. However, the low resolution of these images hampers detection accuracy. To address this, this paper introduces a task-driven super resolution approach that combines super-resolution and...

Full description

Saved in:
Bibliographic Details
Main Authors: Hou, Renkai, Xu, Xiangyang
Format: Conference Proceeding
Language:English
Subjects:
Accuracy
crack segmentation
Image segmentation
Interpolation
Schedules
semantic information
super resolution
Superresolution
Task-driven
Training
Visualization
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Currently, crack images captured by unmanned aerial vehicles are extensively used in the field of crack detection. However, the low resolution of these images hampers detection accuracy. To address this, this paper introduces a task-driven super resolution approach that combines super-resolution and segmentation networks to enhance the segmentation accuracy of low-resolution crack images. Specifically, we designed an end-to-end joint training framework and developed a joint loss function based on super-resolution and segmentation losses, guiding the update of the super-resolution network. This approach not only improves the visual quality of the super-resolved images but also enriches the semantic information in the images. Additionally, the paper explores various joint loss trade-off strategies to determine the optimal implementation of the task-driven method. Experimental results indicate that progressively increasing the proportion of semantic information during training is the most scientific method. Compared to non-task-driven methods, independent training, and BICUBIC interpolation, our approach significantly excels in the segmentation of low-resolution crack images.
ISSN:1934-1768
DOI:10.23919/CCC63176.2024.10662540