Crack segmentation is essential for preventive maintenance in various civil and industrial applications. It makes it possible to identify and divide structural cracks or defects. Complicated sceneries, such as cracks with an irregular form, complicated image environments, and constraints in obtaining global contextual information, affect the performance of crack segmentation. This research proposes an Enhanced-YOLOv8 called YOLOv8-MHSA-TA to reduce the effects of these factors and offer quasi-real-time concurrent identification and segmentation of different crack types. The suggested network uses triplet attention (TA) and multi-head self-attention (MHSA) mechanisms, to enhance YOLOv8’s performance. To evaluate the proposed approach and test its generalization ability, nine public datasets comprising images of civil and industrial structures were collected, including CracK500, Crack3238, Crack Forest Dataset, Deepcrack, Rissbilder, Volker, Sylvie, Magnetic Tile, and Pipeline Gamma Radiography Images. The datasets contain images with cracks of various sizes, shapes, sorts, lighting situations, and orientations. Applying the suggested enhanced YOLOv8 model’s capabilities, cracks are detected and segmented successfully in the examined images. The results demonstrate that, for the Crack500 and Magnetic tile datasets, the suggested model’s segmentation Mean Average Precision (mAP50) is 10.1 and 26.4% higher than that of the original YOLOv8 models. The suggested model was compared with YOLOv8-MHSA, YOLOv8-TA, and the original YOLOv8 models, as well as with other published networks. The outcomes demonstrate that our approach outperforms previously published work and enhances crack segmentation. The outcomes demonstrate that our method outperforms prior published work and enhances crack segmentation when considering the diverse dataset.
Read full abstract