The automatic recognition of design scheme style is a key technology of intelligent design assistance system. Existing methods rely on handcrafted features, and the generalization ability of classifiers is insufficient. This paper proposes a style classification method combining multi-modal feature extraction and improved ensemble classification architecture. At the feature level, a three-modal feature space of structural contour (enhanced histogram of oriented gradients), local texture (encoded by the LBP mutation operator) and global topology (graph convolution embedding) is constructed, and the dimension of the fused feature is compressed to 28% of the original space by a joint PCA‑kernelized dimensionality reduction method, while preserving the nonlinear discriminant structure. At the classification level, a weighted extreme learning machine with adaptive margin adjustment is designed, a multi-branch parallel input and a decision-level weighted voting architecture are introduced, and the loss function is reconstructed by adding a style margin maximization regularization term. 5‑fold stratified cross‑validation is conducted on a dataset containing 5200 samples (8 style types). The experimental results show that the macro F1-score of our method is 88.7%, and the accuracy is 91.5%. Compared with the standard extreme learning machine, the accuracy is improved by 5.3 percentage points, and the generalization error is reduced from 5.8% to 1.2%. Even under strong noise at 5 dB, the accuracy reaches 74.6%, which is 13.8 percentage points higher than the baseline model. The proposed method takes into account both accuracy and robustness, and provides a feasible technical path for efficient recognition of design scheme style.