Abstract: In CAD models, defect recognition arising from surface gaps is essential. Traditional methods, which rely on manual feature extraction and geometric distance thresholds, exhibit notable limitations in multi-scale adaptability and topological discrimination accuracy, making it difficult to meet the demands of complex industrial inspection. To address this issue, a data-driven approach integrating geometric deformation augmentation and deep feature learning is adopted, and a neural network-based gap recognition method using spatial curve convolution operators is proposed. First, a distance-preserving data augmentation algorithm based on radial basis functions was designed. By applying locally rigid transformation constraints, high-quality synthetic samples were generated from the original dataset, alleviating the model generalization bottleneck in small-sample training scenarios. Then, a network incorporating multi-level convolution modules and multi-head self-attention mechanisms was constructed. This network dynamically enhances the geometric salience of gap features and captures long-range dependencies, enabling end-to-end high-precision defect recognition. The original 371 manually constructed samples were expanded to 4 081, which were then randomly divided into training and test sets. The model was trained on a set of 2,860 samples, and after training, performance was evaluated on the remaining test set. Experimental results demonstrate that the proposed method achieves an accuracy of 97.87% and a mean Intersection over Union (mIoU) of 96.42%, with a total testing time of 0.26 s. Its overall performance is significantly superior to that of traditional rule-based methods. This approach provides an efficient and reliable technical solution for intelligent geometric cleaning of CAD models and can be extended to other complex geometric defect recognition scenarios.