esc

Type at least two characters to start searching.

↑↓ select↵ openesc close

Share

Export Citation

APA
MLA
Chicago
Harvard
Vancouver
BIBTEX
RIS
Universitas Hasanuddin
Research output:Contribution to journal›Article›peer-review

Road Damage Detection Using Modified Faster R-CNN with SE-Block and CBAM

Fauzi A.

Proceedings 2026 International Conference on Informationtechnology and Computing Icitcom 2026

Published: 2026

Abstract

Highway surface damage can reduce road safety, service quality, and infrastructure durability if not detected and maintained at an early stage. Conventional road inspection is commonly performed manually, requiring substantial time, labor, and resources, particularly for large-scale road networks. To address this limitation, this study proposes a modified Faster R-CNN framework for UAV-based road surface damage detection. The proposed approach integrates a Squeeze-and-Excitation (SE) Block at C2 and a Convolutional Block Attention Module (CBAM) at C3 of the ResNet-50 backbone to enhance feature representation for small, thin, and low-contrast damage patterns, particularly cracks. Three configurations were evaluated on the independent test set: baseline Faster R-CNN ResNet-50 with FPN, Faster R-CNN with SE at C2 and CBAM at C3, and Faster R-CNN with SE at C2, CBAM at C3, and the default anchor configuration. The attention-enhanced model obtained the highest mAP@50 of 0.689 and recall of 0.792, showing that it was better at detecting road-damage instances. In comparison, the model using the default anchor configuration produced the most balanced performance, achieving the highest mAP50:95 of 0.343, mAP75 of 0.320, precision of 0.532, and F1-score of 0.624. These results indicate that attention-based backbone refinement improves detection sensitivity, while anchor configuration enhances prediction selectivity and localization quality. However, the reduced recall suggests that further optimization is required to improve the detection of small and low-contrast crack objects.

Other files and links

Fingerprint

Environmental scienceSciences
Noise (video)Sciences
Forensic engineeringSciences
Materials scienceSciences
WeldingSciences
LimitingSciences
AccelerationSciences
CalibrationSciences
Compression (physics)Sciences
Feature (linguistics)Sciences