In the realm of structural health monitoring, the automatic detection of cracks on surfaces such as bridges is paramount for ensuring structural integrity. Traditional supervised deep learning models, while capable of predicting the size and location of cracks, are heavily reliant on vast amounts of precisely labeled data, which is not only time-consuming but also impractical in specific scenarios. More critically, these models exhibit limited generalization ability and robustness when confronted with novel or complex datasets. This paper primarily explores the potential of the self-supervised model DinoV2 in crack detection applications, particularly its capability as a powerful visual feature extractor. By leveraging DinoV2 to extract key visual features of crack patterns in images, such as texture, shape and context, we demonstrate how this model can effectively perform crack detection without the support of labeled data. Coupled with a linear classification head, we evaluated the effectiveness of DinoV2 in identifying cracks across a variety of complex backgrounds and compared it with several mainstream supervised learning models. The experimental results indicate that DinoV2 not only enhances the performance of crack detection, especially when dealing with unlabeled images, but also exhibits significant superiority in complex scenarios characterized by high noise and different material textures. This study underscores the application potential of self-supervised models in addressing real-world engineering problems, offering a new perspective for the development of future structural health monitoring technologies.