Accurate roof structure extraction from aerial imagery is important for applications such as solar energy assessment, urban analysis, and 3D city modeling, where roof surfaces require geometrically reliable polygonal representations. Although instance segmentation methods can detect rooftop regions, their masks often contain redundant or overlapping predictions and irregular boundaries, making them difficult to convert into coherent roof surfaces. To address these limitations, we propose a polygon-level structured inference framework for roof-region extraction. Rather than directly using raster masks as final outputs or reconstructing roof polygons from local vector primitives, the proposed method generates over-segmented polygon candidates from line-based cues and aggregates confidence-weighted probabilities at the polygon level. Roof-region assignment is formulated as a Markov Random Field (MRF) optimization problem, where unary terms encode segmentation evidence and pairwise terms enforce spatial consistency. Experiments on the Cities and RoofVec datasets show that the proposed framework suppresses redundant and spurious predictions while improving region coherence. On RoofVec, false positives are reduced from 111 to 94 and precision improves from 96.7% to 97.2% compared with unary-only polygon labeling, while maintaining comparable mean IoU. These results show improved geometric consistency in roof extraction from 2D remote sensing imagery. The source code is publicly available at <code>https://github.com/appleadele/Roof-Structure-Extraction-from-Remote-Sensing-Images</code>.