Surface cracking is one of the most common and informative visible signs of deterioration in masonry and historic structures, and it plays an important role in structural health monitoring and conservation-oriented assessment. However, reliable automatic crack identification remains difficult because masonry surfaces usually contain mortar joints, brick textures, weathering traces, stains, shadows, and uneven illumination, which can be easily mistaken for actual cracks. To address this challenge, this study presents an AI-assisted framework for crack segmentation and geometric feature extraction in masonry structures under complex backgrounds. A pixel-level annotated image dataset is established with three categories, namely background, crack, and mortar, allowing real cracks to be distinguished from masonry texture and joint-related interference. This dataset was used to train the proposed deep learning segmentation model to improve the recognition of fine, irregular, and discontinuous cracks under complex backgrounds. In addition to crack segmentation, the framework extracts quantitative geometric descriptors, including crack length, area ratio, connectivity, and orientation, providing an interpretable representation of visible damage patterns. The proposed approach is intended as a front-end support tool for the inspection, documentation, and conservation of masonry and historic structures. The results indicate that the framework can effectively reduce confusion between cracks and mortar-related features and can provide stable crack identification under strong background interference. The extracted information can support future integration with structural monitoring, digital twins, and numerical models and structural health monitoring systems.