Application of a neural network approach to the recognition of guiding triangles on oil and gas field diagrams

The article is devoted to methods for automated recognition of guiding triangles on engineering diagrams of oil and gas fields. Recognition here refers to the detection, localization, and classification of these graphic objects; the output data are the coordinates and the class of the recognized object. A comparative analysis is carried out for classical filters, the one-stage YOLOv11s-seg detector, and the two-stage Mask R-CNN architecture. A synthetic dataset of 10,000 diagrams was used for training and testing, each of which contains triangles with variations in illumination, overlaps, and incomplete contour rendering. The experiments showed that all the models surpass the classical methods in speed and robustness; however, only Mask R-CNN provides a high level of recognition recall (92 %), whereas YOLO reaches 72 %. The most difficult cases proved to be objects with partial overlaps. It is established that Mask R-CNN demonstrates the greatest robustness to these factors and is recommended for use as part of industrial computer-vision systems for automating the digitization of design documentation.