Open Access

ARTICLE

Enhancing Object Detection in Electrical Substations through Post-Processing Module with Spatial Contexts

Jordán Pascual Espada¹, Lucía Alonso Virgós^2,*, Juan Luis Carús³, Miguel Ángel Fernández Fernández³

1 Faculty of Computer Science and Engineering, Universidad de Oviedo, Oviedo, Asturias, Spain
2 Faculty of Engineering and Technology, Universidad Internacional de la Rioja, Logroño, La Rioja, Spain
3 Faculty of Engineering and Technology, TSK Group, Gijón, Asturias, Spain

* Corresponding Author: Lucía Alonso Virgós. Email:

Computers, Materials & Continua 2026, 89(1), 105 https://doi.org/10.32604/cmc.2026.077447

Received 09 December 2025; Accepted 31 March 2026; Issue published 13 August 2026

Abstract

This research focuses on multi-object detection for interrelated industrial components, such as substation parts. Traditional vision models like YOLO rely primarily on visual features, which limits their ability to: (1) disambiguate visually similar objects using contextual cues like relative position or size, and (2) reinforce low-confidence detections that are spatially plausible given neighboring elements. For instance, a visually similar but semantically incorrect object may be misclassified, or a valid but partially occluded component may be discarded due to a low appearance-based score. Our approach addresses these issues by integrating formalized rules based on relative positions and proportional sizes, enabling contextual reasoning that improves detection accuracy. These challenges are common in industrial inspection scenarios, where components may be occluded, poorly lit, or surrounded by visually similar distractors. Experimental results show that the proposed spatially informed refinement significantly improves precision (from 95.1% to 98.1%) and recall (from 86.5% to 88.9%), leading to a higher F1 score (from 0.905 to 0.933). The system effectively reduces false positives and recovers missed detections, particularly in complex or cluttered scenes, demonstrating the value of integrating spatial logic into object-detection workflows. Although the experiments were conducted on a curated dataset of 714 fixed-camera images from electrical substations, the proposed methodology is scalable and adaptable to other structured domains where spatial relationships are critical for accurate object recognition.

---

Keywords

---