A Multi-Deep Learning Intelligent Surface Rock Crack Identification Method for Transmission Tower Siting

Xiaoxian Tang, Xin Liu, Yuhai Liu, Bowen Zhao, Peng Xie, Jianwen Zhao, Xiaowei Gao, Ran ZhangElectronics 2025 · DOI

TL;DR

This paper introduces a novel "MultiNet" deep learning framework for accurately identifying surface rock cracks, crucial for the safe siting of transmission towers. By combining a CNN for initial image filtering with a UNet-YOLOv8 architecture for precise segmentation, the method significantly improves crack detection accuracy and robustness, outperforming traditional approaches.

What Problem Does This Paper Address?

The safe and stable construction of transmission towers is paramount for reliable power grids. A critical factor in this is the geological assessment of potential tower sites, particularly the identification of surface rock cracks. These cracks can indicate underlying geological instability, leading to foundation settlement, landslides, and even tower collapse. Traditional methods for crack detection, such as manual inspection or basic image processing, are often labor-intensive, inefficient, subjective, and struggle with accuracy and generalization in complex environments. This paper aims to address these limitations by developing an intelligent, automated system that can accurately and reliably identify surface rock cracks.

Methods

The proposed "MultiNet" framework is an end-to-end system that integrates multiple deep learning modules:

  • CNN-Based Image Classification Module: This module acts as a pre-filter. It uses a lightweight Convolutional Neural Network (CNN) to quickly identify and discard images that do not contain any crack features. This significantly speeds up the overall process by focusing subsequent, more computationally intensive steps only on relevant images.
  • UNet-YOLOv8 Crack Segmentation Module: This is the core of the crack identification process. It combines two powerful architectures:
  • UNet: This encoder-decoder network is used for semantic segmentation, providing detailed pixel-level identification of crack features.
  • YOLOv8: This object detection model is employed to generate bounding boxes around potential crack regions. The YOLOv8 bounding boxes are then used as "prior information" to constrain and refine the UNet segmentation results. This integration helps to reduce false positives caused by non-crack features like shadows or potholes, leading to more accurate and robust crack extraction. The paper details the architecture of both UNet and YOLOv8, including their respective down-sampling, up-sampling, and skip connection mechanisms.

  • Optimization Matrix: A Boolean identification matrix (MaskMap) generated by YOLOv8 is used to further refine the UNet's output. This step ensures that the final crack identification is constrained within the detected bounding boxes, effectively removing misidentified features outside these regions.

Datasets

The paper utilizes two distinct datasets for training and evaluation:

  • CNN-Based Crack Classification Module: An open-source surface crack dataset was used. This dataset was manually curated to include images with and without crack features. The training set comprised 4900 images, with a validation set of 10% and a separate test set of 2600 images. Data augmentation techniques were employed to address class imbalance.
  • UNet-YOLOv8 Crack Segmentation Module: A publicly available dataset of 2083 high-resolution (256x256 pixels) crack images with corresponding manually annotated binary masks was used. This dataset features diverse scenarios, including open and closed cracks, high and low-density distributions, and various lighting conditions. It was split into a training set (1666 samples) and a validation set (417 samples) with an 8:2 ratio, and 100 crack feature maps were prepared for testing.

Key Insights and Ideas

  • Synergistic Combination of CNN, UNet, and YOLOv8: The paper demonstrates that combining a CNN for initial filtering with a UNet-YOLOv8 architecture for segmentation offers a powerful and efficient approach to crack identification. The synergy between these models addresses limitations of individual approaches.
  • YOLOv8 as a Constraint for UNet: Using YOLOv8's bounding box predictions to guide and constrain UNet's segmentation is a key innovation. This effectively mitigates false positives from shadows and other non-crack features, significantly improving accuracy.
  • End-to-End System for Practical Application: The development of a complete, end-to-end system that moves from raw image input to optimized crack identification is crucial for practical engineering applications, reducing the need for manual intervention and complex feature engineering.
  • Generalization Across Scenarios: The method's ability to perform well on various crack types (asphalt, concrete, wall cracks) and its strong performance on different datasets highlight its adaptability and potential for broader use.

Key Results

  • CNN Module Performance: The CNN-based classification module achieved a testing accuracy of 93.23%, demonstrating its effectiveness in filtering out irrelevant images.
  • UNet-YOLOv8 Module Performance: The combined UNet-YOLOv8 module achieved high evaluation metrics on the rock crack dataset, with Precision, Recall, Accuracy, and Specificity converging to 79%, 70%, 97.5%, and 98.9% respectively during training. The test set accuracy for the overall MultiNet method was reported as 96.41%.
  • Comparison with Other Methods: The proposed method outperformed traditional machine learning algorithms like SVM, Random Forest, Naive Bayes, and Decision Tree across key metrics (Accuracy, Recall, F1-value, mAP).
  • Qualitative Improvements: Visual results show that the optimized method significantly reduces interference from shadows and other non-crack features compared to unoptimized approaches.

Limitations and Future Work

  • Dataset Specificity: While the paper tests on various crack types, the primary focus and extensive validation are on rock cracks for transmission tower siting. Further validation on a wider range of infrastructure and geological conditions could be beneficial.
  • Computational Cost: Although the CNN filter improves efficiency, deep learning models can still be computationally intensive. Optimizing for real-time deployment on edge devices might be a future direction.
  • Further Generalization: While the paper shows good generalization, exploring its performance on extremely challenging conditions (e.g., very low light, occlusions, complex textures) could be a valuable next step.
  • Integration with Other Data Sources: Future work could explore integrating this visual crack identification with other geological data (e.g., sensor data, ground penetrating radar) for a more comprehensive site assessment.

Summary updated

Abstract

Accurate identification of surface cracks is of great significance for the site selection of transmission towers, as it directly affects the safety and stability of power grid construction. Traditional manual inspection methods are labor-intensive and inefficient, making them inadequate for large-scale and high-precision applications. Consequently, intelligent crack recognition technologies are receiving increasing attention and adoption. This study proposes a novel surface crack identification method that integrates multiple neural networks, aiming to overcome the limitations of traditional crack identification approaches, such as low accuracy and poor generalization. The proposed framework incorporates a convolutional neural network (CNN)-based data filtering module and a crack segmentation module combining UNet and YOLOv8 architectures. Together, these components form a robust and accurate end-to-end crack identification system, which is further applied to the evaluation of rock fragmentation. To verify the effectiveness and accuracy of the proposed method, experiments were conducted on a rock crack dataset and compared against several existing approaches. The results demonstrate that the proposed method achieves superior performance in crack detection accuracy. Moreover, tests on various scenario datasets also yielded promising results, indicating the strong generalization ability and adaptability of the method.
BibTeX
@inproceedings{tang2025multi,
  author = {Xiaoxian Tang and Xin Liu and Yuhai Liu and Bowen Zhao and Peng Xie and Jianwen Zhao and Xiaowei Gao and Ran Zhang},
  title = {{A Multi-Deep Learning Intelligent Surface Rock Crack Identification Method for Transmission Tower Siting}},
  booktitle = {Electronics},
  year = {2025},
  doi = {10.3390/electronics14112255},
}
Figure