In view of the bottleneck problems in the creation and interpretation of traditional engineering drawings, including poor compatibility across formats, low accuracy in symbol recognition, and insufficient adaptability to standardization, this study proposes an intelligent analysis and adaptive error-correction framework based on deep visual cognition. First, a geometric dimensioning and tolerancing (GD&T) annotation knowledge graph is constructed in compliance with the American society of mechanical engineers (ASME) Y14.5–2018 standard, resolving semantic ambiguities. The intelligent adaptation of drawing formats is realized through multi-modal feature fusion technology.Subsequently, a detection-segmentation collaborative architecture is developed by integrating a task dynamic alignment detection head (TDADH) and a high frequency and spatial perception feature pyramid network (HS-FPN) for tiny object detection. Additionally, efficient local attention (ELA) for deep convolutional neural networks is embedded, and cross-modal transfer learning strategies are designed. Finally, an intelligent analysis engine with self-correction capabilities is implemented to achieve a dimension tolerance identification error rate of less than 0.8%. The experimental results demonstrate a standardized conversion accuracy of 98.6%, and the framework enables seamless integration with mainstream computer aided design (CAD) platforms for 3D model reconstruction.