Abstract
Maintaining national power grids requires frequent aerial inspections of millions of components. A critical challenge in this process is object re-identification (Re-ID)—the ability to match specific assets, such as electric poles, across different inspection flights taken from varying angles and altitudes. This paper presents a hybrid Siamese neural network that combines convolutional visual features with structural graph data. By modeling electric poles as graphs of constituent parts (e.g., insulators, transformers), the system achieves a CMC@1 of 0.9306, significantly outperforming traditional image-only methods.
Introduction
In Italy alone, the national power grid spans over 74,000 km, requiring constant monitoring via helicopters and drones. For engineers, the primary difficulty lies in identifying whether an asset in a current image is the same individual unit recorded in previous surveys. Unlike person Re-ID, where clothing provides high-variance cues, industrial assets like electric poles are often structurally identical and set against similar rural backgrounds. Furthermore, changes in camera perspective, lighting, and weather can drastically alter the global visual signature of a pole, making standard Convolutional Neural Network (CNN) approaches insufficient.
Methodology: The Hybrid Pipeline
The researchers propose that the unique "fingerprint" of an industrial asset lies in its topology—the specific spatial arrangement of its sub-components. To capture this, they developed a multi-stage hybrid pipeline.
- Asset Detection: An EfficientDet-D2 model identifies the pole and its individual components (e.g., jumpers, insulators).
- Visual Branch: Detected assets are cropped and processed through a ResNet-34 to extract texture and appearance embeddings.
- Graph Branch: The system constructs a graph where nodes represent assets and edges represent spatial proximity. To process this structural data, the authors utilized Chebyshev Graph Neural Networks, which use higher-order polynomials to aggregate neighborhood information more effectively than standard GCNs.
- Feature Fusion: The visual and structural embeddings are concatenated to form a comprehensive representation of the asset.
Training Strategy: Metric Learning
To train the model to distinguish between identical-looking poles, the authors employed a Siamese architecture governed by a triplet loss function. This approach compares three samples simultaneously: an anchor image, a positive sample (the same pole from a different angle), and a negative sample (a different but similar-looking pole).
The triplet loss forces the network to map the anchor and positive samples close together in the embedding space while pushing the negative sample away. This ensures that the model learns to prioritize the subtle structural differences that define a unique asset.
Experimental Results
The study utilized a proprietary dataset from Enel S.p.A., featuring nearly 9,000 images of electrical lines with diverse morphologies. The hybrid model demonstrated superior performance across all key metrics:
- Top-1 Accuracy (CMC@1): 0.9306
- Mean Average Precision (mAP): 0.788
Comparatively, a standard Siamese CNN focusing only on the pole's global appearance achieved a CMC@1 of 0.876. The inclusion of graph-based structural information provided a 6 percentage point relative improvement, proving that topology is a more stable descriptor than raw visual appearance in aerial Re-ID tasks.
Conclusion
By fusing visual deep learning with Graph Neural Networks, this research provides a robust framework for industrial object re-identification. The results suggest that structural reasoning—understanding how parts relate to one another—is essential for distinguishing between complex, mass-produced objects in cluttered environments. Future work may involve end-to-end training of the detector and Re-ID components to further enhance system robustness.
Summary updated