HCDFI-YOLOv8: A Transmission Line Ice Cover Detection Model Based on Improved YOLOv8 in Complex Environmental Contexts

Lipeng Kang, Feng Xing, Tao Zhong, Caiyan QinSensors 2025 · DOI

TL;DR

This paper introduces HCDFI-YOLOv8, an enhanced object detection model designed to accurately identify ice cover on transmission lines, even in challenging environmental conditions. By introducing novel modules for feature extraction, fusion, and loss calculation, the model significantly improves detection accuracy compared to standard YOLOv8 and other state-of-the-art methods.

What Problem Does This Paper Address?

Detecting ice accumulation on transmission lines is crucial for maintaining a stable power grid, especially with the increasing frequency of extreme weather events. However, current methods, particularly those using Unmanned Aerial Vehicles (UAVs), struggle with accuracy due to variable shooting angles, complex backgrounds, and the inherent difficulty in distinguishing ice from other environmental features. This leads to issues like false detections and missed ice coverage, posing a risk to power infrastructure. The paper aims to improve the robustness and accuracy of these detection systems.

Methods

The HCDFI-YOLOv8 model is built upon the YOLOv8 architecture but incorporates several key innovations:

  • Cross-Dense Hybrid (CDH) Module: This novel module is integrated into the backbone network. It uses parallel heterogeneous convolutional kernels to process features, aiming to improve detection accuracy while mitigating the increase in computational load (floating-point operations) that often accompanies model enhancements.
  • Improved C2f_DFF Module: In the neck network, the standard C2f module is replaced with an improved C2f_DFF (CSPDarknet53 to 2-Stage FPN_Dynamic Feature Fusion) module. This aims to reduce feature loss during the fusion of shallow and deep network features, enhancing the model's ability to retain crucial local information.
  • Feature Adaptive Spatial Feature Fusion (FASFF) Detection Head: The detection head is optimized with the FASFF module. This module employs an adaptive spatial feature fusion strategy to better extract features across different scales, crucial for detecting ice of varying sizes and appearances.
  • Inner-Complete Intersection over Union (Inner_CIoU) Loss Function: A new loss function, Inner_CIoU, is introduced to address limitations of the original CIoU loss in YOLOv8. It uses auxiliary bounding boxes of different scales to improve regression accuracy and accelerate model convergence.

Datasets

The study utilizes a public dataset from CSDN containing 1372 images of transmission line icing. This dataset includes two categories: "mixed rime" (opaque ice and snow) and "glaze" (transparent ice). The dataset was split into 1144 images for training and 228 for validation. The authors note that this dataset has diverse icing types, sizes, and background environments, making it suitable for enhancing model robustness.

Key Insights and Ideas

  • Balancing Accuracy and Efficiency: A core challenge in improving object detection models is the trade-off between increased accuracy and computational cost. The CDH module specifically targets this by using smaller kernels strategically, suggesting that clever architectural design can boost performance without a prohibitive increase in FLOPS.
  • Multi-scale Feature Fusion is Key: Transmission line icing can appear at various scales due to distance and ice thickness. The FASFF head's ability to adaptively fuse features from different levels is highlighted as critical for comprehensive detection.
  • Beyond Standard IoU: The limitations of standard IoU-based loss functions are well-known. The introduction of Inner_CIoU, which considers inscribed rectangles and scales, demonstrates a practical approach to refining bounding box regression and improving convergence.
  • Synergistic Module Effects: The paper emphasizes that the proposed modules not only improve performance individually but also exhibit synergistic effects when combined, leading to greater gains than the sum of their parts.

Key Results

The HCDFI-YOLOv8 model demonstrated significant improvements:

  • Accuracy: It achieved a 2.7% improvement in [email protected] and a 2.5% improvement in [email protected]:0.95 compared to the standard YOLOv8 model.
  • Comparative Performance: Among twelve models evaluated for icing detection, HCDFI-YOLOv8 delivered the highest overall detection accuracy.
  • Specific Gains: When compared to other established models like TOOD and Faster R-CNN, HCDFI-YOLOv8 showed notable increases in precision, recall, and [email protected], while often reducing FLOPs. For instance, it achieved a 90.2% [email protected] and 70.6% [email protected]:0.95, outperforming YOLOv8's 87.5% and 68.1% respectively.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Simplicity: The current dataset, while diverse, may have relatively simple backgrounds. Performance might be affected in more complex weather conditions or terrains.
  • Aspect Ratio Dependence: Transmission line objects often have large aspect ratios, making the model's reliance on high-quality labeled data critical for training.
  • Future Directions:
    • Fusion Network: Building a fusion network that integrates "image + meteorological time series data + 3D terrain" to account for spatial and environmental factors more robustly.
  • Enhanced Dataset: Utilizing Oriented Bounding Boxes (OBB) to create a more accurate ice cover dataset, further improving model precision.

Summary updated

Abstract

When unmanned aerial vehicles (UAVs) perform transmission line ice cover detection, it is often due to the variable shooting angle and complex background environment, which leads to difficulties such as poor ice-covering recognition accuracy and difficulty in accurately identifying the target. To address these issues, this study proposes an improved icing detection model based on HCDFI-You Only Look Once version 8 (HCDFI-YOLOv8). First, a cross-dense hybrid (CDH) parallel heterogeneous convolutional module is proposed, which can not only improve the detection accuracy of the model, but also effectively alleviate the problem of the surge in the number of floating-point operations during the improvement of the model. Second, deep and shallow feature weighted fusion using improved CSPDarknet53 to 2-Stage FPN_Dynamic Feature Fusion (C2f_DFF) module is proposed to reduce feature loss in neck networks. Third, optimization of the detection head using the feature adaptive spatial feature fusion (FASFF) detection head module is performed to enhance the model's ability to extract features at different scales. Finally, a new inner-complete intersection over union (Inner_CIoU) loss function is introduced to solve the contradiction of the CIOU loss function used in the original YOLOv8. Experimental results demonstrate that the proposed HCDFI-YOLOv8 model achieves a 2.7% improvement in [email protected] and a 2.5% improvement in [email protected]:0.95 compared to standard YOLOv8. Among twelve models for icing detection, the proposed model delivers the highest overall detection accuracy. The accuracy of the HCDFI-YOLOv8 model in detecting complex transmission line environments is verified and effective technical support is provided for transmission line ice cover detection.
BibTeX
@inproceedings{kang2025hcdfi,
  author = {Lipeng Kang and Feng Xing and Tao Zhong and Caiyan Qin},
  title = {{HCDFI-YOLOv8: A Transmission Line Ice Cover Detection Model Based on Improved YOLOv8 in Complex Environmental Contexts}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25175421},
}
Figure