Traditional manual crack inspection methods suffer from low efficiency and accuracy that is highly dependent on subjective judgment. In contrast, although deep learning–based crack detection techniques have attracted considerable attention in recent years, their performance on speckle images remains limited due to weak sensitivity to crack features, resulting in low recognition accuracy. To address this issue, this paper proposes a crack detection approach based on supervised fine-tuning, in which a pretrained Attention U-Net model is specifically trained on a speckle crack dataset to effectively transfer crack recognition knowledge to speckle-image scenarios. Experimental results demonstrate that, compared with models pretrained on open datasets, the proposed method improves the Intersection over Union (IoU) by 27.4% and the Pixel Accuracy (PA) by 20.4%. In addition, the crack length error obtained using the proposed method is within 5%, while the crack width error is within 10%. These results indicate that the proposed approach can significantly enhance the accuracy and automation level of crack detection in speckle images, offering a promising solution for structural health monitoring and safety assessment.