Multi-scale feature fusion for pavement crack detection based on Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Automated pavement crack detection is challenging due to slender crack identification and varying lighting.
• Existing methods struggle with complex backgrounds and illumination inconsistencies.

Purpose of the Study:

• To propose a novel multi-scale feature fusion network based on the Transformer architecture for improved automated pavement crack detection.
• To enhance feature extraction and context acquisition for more accurate crack segmentation.

Main Methods:

• Utilized a Transformer architecture in the encoding phase for global modeling and long-distance dependence.
• Employed dilated convolution to expand the receptive field while preserving feature map resolution.
• Implemented multi-scale feature fusion in the decoding phase to restore detailed crack information.

Main Results:

• Achieved an F1 score of 70.84% on the Crack500 dataset, a 1.42% improvement over the state-of-the-art.
• Achieved an F1 score of 84.50% on the DeepCrack dataset, a 2.07% improvement over the state-of-the-art.
• Demonstrated superior detection accuracy and generalization across high and low brightness conditions.

Conclusions:

• The proposed Transformer-based multi-scale feature fusion network significantly improves automated pavement crack detection accuracy.
• The method exhibits robust performance under challenging lighting conditions, outperforming existing state-of-the-art approaches.
• The approach offers better generalization capabilities for pavement crack analysis.