In today’s rapidly advancing digital era, 3D building models play a crucial role in the automation of architectural design and construction processes. However, traditional door modeling methods, which rely on manual identification and modeling, require substantial time and manual effort. Therefore, employing image recognition technology for the automated identification of doors is essential. Among image recognition technologies, although YOLO v8 excels in object detection, it still has shortcomings in recognizing small targets and processing complex images, such as door identification. To improve performance, three key enhancements were introduced to the YOLO v8 model: (1) Optimization of image preprocessing to increase efficiency and accuracy of the decision layer.(2)A door recognition dataset containing 15,663 images was utilized, ensuring extensive material and diverse samples.(3)Parameter optimization was conducted, including the adjustment of the learning rate and training epochs. The improved YOLO v8 model achieved an accuracy of approximately 90% in identifying doors in complex architectural floor plans, as demonstrated by experimental findings.Compared to HRNetsV2, CNN-GCN, and DeepLabv3, the precision of door detection has increased by 5%.