Multi-resolution adaptive attention fusion network for automated damage detection and quantification on bicycle roads

The growing adoption of bicycles as an eco-friendly mode of transportation has increased the demand for efficient maintenance systems for bicycle roads. This research proposes a novel multi-resolution adaptive attention fusion (MAAF) network for automated damage detection and quantification. Conventional deep learning networks often face challenges due to vision image resolution dependency from varying sizes damage detection with fixed camera pixel. In addition, it is struggled to accurately identify damage within diverse surface colors and textures to bicycle pavements. To resolve these technical issues, the proposed MAAF network incorporates color-based damage classification and negative sample class training to minimize false detections by the bicycle road pavements' characteristics. Then, computer vision-based image processing algorithms are utilized for damage quantification. This comprehensive framework establishes an automated maintenance process specifically optimized for bicycle road characteristics. Experimental validation on actual roads demonstrated F1-scores of 0.903 (cracks), 0.919 (alligator cracks), 0.878 (scaling), and 0.957 (potholes).