Hsod: hybrid strategy object detection in piping and instrumentation diagrams and process flow diagrams

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As core technical documents for the process industry, Piping and Instrumentation Diagrams (P&IDs) and Process Flow Diagrams (PFDs) face severe challenges in automated object detection: small industrial symbols (e.g., arrows, stream numbers, valves) occupy minimal image areas, leading to insufficient feature extraction; limited labeled samples result in poor model generalization under few-shot conditions; and dense symbol distribution with complex background interference causes high false detection and missed detection rates. To address these issues, this study constructs a multi-industry P&ID–PFD dataset based on 182 labeled original industrial flow diagrams. The dataset contains over 14,000 sliced image patches and approximately 160,000 annotated object instances across 308 core object categories, covering representative scenarios from the petrochemical and coal chemical industries. Furthermore, we propose a Hybrid Strategy Object Detection (HSOD) workflow that integrates parallel full-image and overlapping-patch detection and hybrid post-processing. Based on YOLOv12, the workflow employs SAHI with 640 × 640 patches and 50% overlap to enhance small object feature extraction, and designs a hybrid filtering strategy combining position-based confidence calibration, adaptive non-maximum suppression (NMS), and quartile-based confidence thresholding to eliminate duplicate detections and false positives. Extensive experiments show that our HSOD workflow outperforms mainstream baselines (RT-DETR, YOLOv12, Faster R-CNN) on all metrics. Relative to the strongest baseline for each metric, HSOD improves mAP@0.5 by 23.27 percentage points over RT-DETR and mAP@0.5:0.95 by 24.41 percentage points over Faster R-CNN. At a confidence threshold of 0.5, HSOD achieves a small-object recall of 0.9774, exceeding Faster R-CNN by 36.61 percentage points. Qualitative comparison with Gemini 3.1 Pro further validates the superiority of specialized detection models. This work provides a reliable solution for the intelligent digital parsing of industrial drawings, supporting the intelligent operation, maintenance and digital twin construction of process industries.