Evaluating YOLO models in Power Lines Detection using TTPLA Dataset

Abstract

Automated inspection of power grid infrastructure, including wires and transmission towers, is essential for ensuring the safety and reliability of electricity distribution systems. This study evaluates the performance of state-of-the-art object detection algorithms from the YOLO (You Only Look Once) family (YOLOv5, YOLOv8, and YOLOv11) using the Transmission Tower and Power Line Aerial-Image (TTPLA) dataset. The TTPLA dataset, characterized by high-resolution aerial imagery with annotated bounding boxes, presents challenges such as complex backgrounds, diverse object geometries, and class imbalance. Among the models tested, YOLOv8 achieved the highest precision and recall metrics, demonstrating exceptional accuracy in detecting wires and transmission towers under real-world conditions. By leveraging advanced deep learning techniques, addressing class imbalance, and fine-tuning hyperparameters, this work establishes a new benchmark for UAV-based power grid inspections, highlighting the potential of cutting-edge computer vision models in infrastructure monitoring.

System model

The architecture of our solution is described in the figure below and is divided into four main steps: data acquisition (TTPLA), data preparation, labels creation, models training, and evaluation.

Algorithms used

• YOLOv5:
  This version is lightweight and fast, making it highly suitable for real-time applications such as wire detection. It benefits from extensive community support and a wide range of pre-trained models for easy customization. However, its accuracy is slightly lower compared to newer versions, particularly for detecting thin and intricate objects like wires. It strikes a good balance between speed and performance, focusing on simplicity and ease of deployment.

• YOLOv8:
  Building on its predecessor, YOLOv8 features an improved model architecture and greater accuracy, making it better equipped to detect small or intricate objects such as wires in cluttered environments. It incorporates enhanced backbone networks and optimized training pipelines. Despite these improvements, it demands higher computational resources, which might limit its use on devices with constrained hardware capabilities.

• YOLOv11:
  As the latest iteration, YOLOv11 offers state-of-the-art performance with advanced detection capabilities for complex scenarios, such as detecting dense or overlapping wires. It integrates cutting-edge techniques like transformer-based architectures, improving its ability to capture fine details. However, it requires significant computational power and longer training times, and it has less community support compared to earlier versions, making it a less accessible option for some applications.

Running the code

Results

The main results are illustrated in the table and figure below:

Performance Evaluation of the YOLO Models Compared with Existing Methods

Model	Aspect Ratio	mAP50%	mAP50-95%	Precision (B)	Recall (B)	Fitness
ResNet-50	700x700	42.62	21.90	-	-	-
	550x550	43.37	20.76
	640x360	46.72	16.50
ResNet-101	700x700	43.19	22.96	-	-	-
	550x550	45.30	22.61
	640x360	44.99	18.42
YOLOv5	700x700	48.19	33.37	63.19	48.51	34.85
	550x550	45.34	30.79	59.13	46.63	32.25
	640x360	45.54	31.09	63.41	44.49	32.53
YOLOv8	700x700	48.23	34.24	64.61	48.26	35.63
	550x550	45.93	31.66	60.51	47.33	33.09
	640x360	45.45	31.25	64.93	44.43	32.67
YOLOv11	700x700	46.80	32.39	64.30	47.33	33.83
	550x550	44.93	30.91	60.84	45.86	32.31
	640x360	44.19	29.89	61.49	44.09	31.32

Citation

This work comes as a Work In Progress contribution to the dataset paper TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines of Dr. Abdelfattah et al.

@inproceedings{abdelfattah2020ttpla,
  title={TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines},
  author={Abdelfattah, Rabab and Wang, Xiaofeng and Wang, Song},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2020}
}

If you find this project interesting, please do not hesitate to reach out to me for any recommendations, questions, or suggestions. Email me