On Salient Concrete Crack Detection Via Improved Yolov5

Conventional visual inspection of concrete structures in Structural Health Monitoring (SHM) ascertains the health status of concrete structures but due to subjectivity and intense labor on the part of experts, the dynamics have almost shifted to automation which heavily dwells on detectors. Although object detectors accelerate the task of SHM, in concrete crack detection, some challenges include the asymmetrical shapes, hairline/salient cracks, and varying scales of cracks. To mitigate these challenges, we introduce image denoising and attention mechanisms as enhancements for a one-stage detector. An extensive experiment on the SDNET2018 dataset indicates improvement on the adapted detector which attains a little over 2% mAP than two baseline detectors.