Abstract: Rapid assessment of building damage under blast loading is a critical component of strike effect evaluation and battlefield situation awareness. An intelligent blast damage assessment method based on close-range image semantic segmentation was proposed. First, the correspondence between blast-induced structural damage mechanisms and image visual features was analyzed, and a labeled dataset containing 163 high-resolution close-range images with 8 semantic categories was constructed for battlefield building blast damage. Second, quantitative assessment metrics including damage area ratio, crack density, and damage grade were defined based on segmentation results. On this basis, three mainstream semantic segmentation models—DeepLabV3+, SegFormer, and Mask2Former—were systematically compared in terms of segmentation performance and damage assessment capability. To ensure statistical reliability, the experiment was conducted with image-level data splitting and repeated trials using three random seeds. Results show that Mask2Former achieves the best segmentation accuracy, outperforming SegFormer and DeepLabV3+ by 4.90 and 7.60 percentage points respectively. The performance gap is primarily manifested in rare damage categories, where the object-level loss mechanism of Mask2Former demonstrates significant advantages over pixel-level loss models in crack and person segmentation. Crack remains the most challenging category, and its missed detection leads to severe underestimation of crack density, thereby affecting damage grade determination. The impact of segmentation errors on damage assessment exhibits significant asymmetry: severe damage can be reliably determined, while mild and moderate damage grades are highly sensitive to the recognition accuracy of cracks and breaches. This study provides a data foundation, methodological framework, and model selection reference for rapid intelligent assessment of battlefield building blast damage.