To improve the flattening leveling quality and efficiency for heavy plates and achieve intelligent flattening process, machine vision is used instead of manual recognition to achieve accurate recognition of the contour of warped plates in this article, and modeling research is conducted on the relevant process parameters of point cloud recognition and flattening process. A structured light camera was applied to obtain the point cloud data of the three-dimensional profile of warped heavy plates, and the point cloud data was preprocessed, such as denoising, segmentation and simplification. The three-dimensional surface of the warped plate was reconstructed by least square method. According to the three-point bending leveling theory and surface theory in differential geometry, the curvature of three-dimensional warped plate was calculated. Not only the fulcrum positions were optimized, but also the screw-down force and displacement were calculated, realizing the modeling for the intelligent flattening leveling of heavy plate. It was further verified by finite element method and experiments. The comparison between theories and experiments shows that the reconstructed three-dimensional surface is consistent with the actual warped heavy plate. The deviation between the calculated screw-down force based on the reconstructed three-dimensional surface and the measured force is about 2.21%, and the initial unevenness of 17.2 mm.m-1 is reduced to 3.28 mm.m-1, indicating the modelling is relatively accurate. This method is feasible and provides a theoretical model for intelligent flattening leveling of heavy plates.