Automated Detection for Concrete Surface Cracks Based on Deeplabv3+ BDF

Concrete cracks have always been the focus of research because of the serious damage they cause to structures. With the updating of hardware and algorithms, the detection of concrete structure surface cracks based on computer vision has received extensive attention. This paper proposes an improved algorithm based on the open-source model Deeplabv3+ and names it Deeplabv3+ BDF according to the optimization strategy used. Deeplabv3+ BDF first replaces the original backbone Xception with MobileNetv2 and further replaces all standard convolutions with depthwise separable convolutions (DSC) to achieve a light weight. The feature map of a shallow convolution layer is additionally fused to improve the detail segmentation effect. A new strategy is proposed, which is different from the two-stage training. The model training is carried out in the order of transfer learning, coarse-annotation training and fine-annotation training. The comparative test results show that Deeplabv3+ BDF showed good performance in the validation set and achieved the highest mIoU and detection efficiency, reaching real-time and accurate detection.

Cette oeuvre a été publiée sous la license Creative Commons Attribution 4.0 (CC-BY 4.0). Il est autorisé de partager et adapter l'oeuvre tant que l'auteur est crédité et la license est indiquée (avec le lien ci-dessus). Vous devez aussi indiquer si des changements on été fait vis-à-vis de l'original.