Abstract

The validation of technical drawings and blueprints is a critical step in engineering projects, yet it remains a largely manual and time-consuming process. This thesis investigates the use of deep learning-based object detection to automate the identification and classification of sprinkler symbols in technical blueprint drawings, as a first step toward automating the validation process at the Swedish engineering consultancy Incoord. Two object detection architectures were evaluated and compared: YOLO, a single-stage detector implemented via the Ultralytics framework, and Faster R-CNN, implemented via Detectron2. Both models were fine-tuned on a dataset of 100 sliced blueprint images containing 981 annotated sprinkler symbols across six classes. Image slicing into 1024×1024 pixel patches was found to be a prerequisite for meaningful detection performance, as both models produced near-zero results when trained on full-resolution images. YOLO11m outperformed all Faster R-CNN variants across all evaluation metrics, achieving a validation mAP50 of 0.989 in 25 minutes of training. The performance gap is attributed to YOLO’s more effective convergence on small datasets, and its built-in data augmentation pipeline. An inference evaluation on one seen and one unseen blueprint using Slicing Aided Hyper Inference (SAHI) showed that the model generalizes well to unseen drawings for well-represented classes, while performance on minority classes with few training examples remained a limiting factor. A proof-of-concept inference tool was developed that runs SAHI-based detection on full-resolution blueprints and visualizes detections with coverage radius overlays and spacing violation warnings. The results demonstrate that machine learning can meaningfully extract spatial information from sprinkler blueprints for validation purposes. Generalization across multiple projects remains an open challenge that future work should address through expanded training data and project-specific fine-tuning strategies.