Improved Algorithm for Insulator and Its Defect Detection Based on YOLOX

Gujing Han, Tao Li, Qiang Li, Feng Zhao, Min Zhang, Ruijie Wang, Qiwei Yuan, Kaipei Liu, Liang QinSensors 2022 · DOI

TL;DR

This paper introduces an improved YOLOX-S model for detecting insulators and their defects in aerial images. By incorporating a novel SIoU-d regression loss function and an Efficient Channel Attention (ECA) module, the model achieves significantly higher accuracy (97.18% mAP) and maintains a fast detection speed (71 fps), addressing challenges posed by small targets and complex backgrounds.

What Problem Does This Paper Address?

Detecting defects on aerial insulators is crucial for maintaining the stability of high-voltage transmission lines. However, existing methods struggle with several challenges:

  • Small Targets: Insulator defects often appear as very small objects in aerial images, making them difficult to detect accurately.
  • Complex Backgrounds: The natural environment surrounding transmission lines can be cluttered and visually similar to defects, leading to false positives or missed detections.
  • Accuracy vs. Speed Trade-off: While some models achieve high accuracy, they might be too slow for real-time inspection. Conversely, faster models may sacrifice accuracy, especially for small or subtle defects.
  • Regression Loss Limitations: Standard IoU-based loss functions can be insensitive to the orientation and precise positioning of small bounding boxes, hindering accurate regression and model convergence.

This paper aims to overcome these limitations by enhancing the YOLOX-S object detection model to achieve both high accuracy and real-time performance for insulator defect detection.

Methods

The researchers propose two key improvements to the YOLOX-S architecture:

  1. Improved Regression Loss (SIoU-d):
    • The standard YOLOX-S uses IoU loss, which has limitations in guiding the regression of small bounding boxes, especially when they don't intersect or are in different positions within the ground truth.
  2. The paper introduces SIoU-d (Scale-Invariant IoU with Directional angle), an enhanced loss function. SIoU-d incorporates:
  3. Centroid Distance and Shape Information: It considers the distance between box centroids and their relative widths/heights, making it scale-invariant.
  4. Directional Angle Loss: Crucially, it adds an angle loss that encourages the predicted box's center to regress towards the ground truth's center along the diagonal direction of the ground-truth box. This is inspired by the observation that diagonal convergence can accelerate learning and improve accuracy.
  5. This improved loss function aims to accelerate model convergence and achieve better regression accuracy, particularly for small targets.
  6. Attention Mechanism (ECA):
    • To combat the issue of complex backgrounds and redundant features, the researchers embed the Efficient Channel Attention (ECA) module.
  7. ECA is placed between the backbone (feature extractor) and the feature fusion layers (Neck).
  8. ECA adaptively assigns weights to feature channels, allowing the model to focus on more important features and suppress irrelevant ones from the complex background. This helps improve the model's sensitivity to subtle defects.
  9. The paper also explored other attention mechanisms (SE, CBAM, Non-Local) but found ECA to offer a good balance between performance improvement and computational overhead, maintaining high detection speed.

Additionally, the paper mentions replacing SPP (Spatial Pyramid Pooling) with SPPF (Spatial Pyramid Pooling Fast) for slight computational efficiency gains and replacing PANet with FPN (Feature Pyramid Network) for feature fusion.

Datasets

The study utilized a dataset of 1588 aerial images of insulators, containing 2908 insulators in total, with 715 of them exhibiting defects. The dataset was split into training (1286 images), validation (143 images), and testing (159 images) sets with an 8:1:1 ratio. The paper does not mention introducing a new benchmark dataset but uses this curated dataset for their experiments.

Key Insights and Ideas

  • Diagonal Regression is Key for Small Targets: The paper's most significant insight is that forcing bounding box regression along the diagonal of the ground-truth box, rather than just axis-aligned, can dramatically improve convergence speed and accuracy for small targets. This is because it more effectively guides the center point and shape alignment.
  • Attention for Background Suppression: Integrating channel attention mechanisms like ECA is highly effective in filtering out noise and irrelevant features from complex backgrounds, allowing the model to better focus on subtle insulator defects.
  • SIoU-d's Superiority: The proposed SIoU-d loss function demonstrably outperforms standard IoU loss by addressing its shortcomings related to non-overlapping boxes and orientation sensitivity.
  • YOLOX-S as a Strong Baseline: YOLOX-S, even before improvements, is recognized for its speed. The enhancements build upon this strong foundation to achieve state-of-the-art performance for this specific task.

Key Results

The improved YOLOX-S model achieved impressive results:

  • Overall Accuracy: A mean Average Precision (mAP) of 97.18%, which is a 2.74% increase compared to the baseline YOLOX-S.
  • Defect Detection: Achieved 97.79% AP for insulator defects.
  • Insulator Detection: Achieved 96.57% AP for insulators.
  • Detection Speed: Maintained a fast detection speed of 71 frames per second (fps).
  • Comparison to Other Models: Algorithm 4 (which incorporates FPN, SIoU-d, and ECA) outperformed other established models like Faster-RCNN, SSD, YOLOv3, YOLOv4, and YOLOv5-S in terms of AP for both insulators and defects.

Qualitatively, the improved model demonstrated a superior ability to detect insulators in complex backgrounds and accurately identify defects that were missed or poorly localized by the original YOLOX-S model.

Limitations and Future Work

  • Limitations:
    • While the paper claims the improved model can detect insulators and defects "accurately and in real-time," the term "accurately" is relative. The paper focuses on improving existing metrics rather than achieving perfect detection.
  • The study primarily focuses on aerial images. The performance on ground-level or different types of insulator defects might vary.
  • The computational cost of attention mechanisms, while managed with ECA, can still be a factor for extremely resource-constrained environments.
  • Future Work:
    • The paper implicitly suggests further exploration of attention mechanisms and regression loss functions.
  • Expanding the dataset to include more diverse scenarios, weather conditions, and defect types would enhance the model's generalization capabilities.
  • Investigating real-time deployment on edge devices for practical inspection applications.

Summary updated

Abstract

Aerial insulator defect images have some features. For instance, the complex background and small target of defects would make it difficult to detect insulator defects quickly and accurately. To solve the problem of low accuracy of insulator defect detection, this paper concerns the shortcomings of IoU and the sensitivity of small targets to the model regression accuracy. An improved SIoU loss function was proposed based on the regular influence of regression direction on the accuracy. This loss function can accelerate the convergence of the model and make it achieve better results in regressions. For complex backgrounds, ECA (Efficient Channel Attention Module) is embedded between the backbone and the feature fusion layer of the model to reduce the influence of redundant features on the detection accuracy and make progress in the aspect. As a result, these experiments show that the improved model achieved 97.18% mAP which is 2.74% higher than before, and the detection speed could reach 71 fps. To some extent, it can detect insulator and its defects accurately and in real-time.
BibTeX
@inproceedings{han2022improved,
  author = {Gujing Han and Tao Li and Qiang Li and Feng Zhao and Min Zhang and Ruijie Wang and Qiwei Yuan and Kaipei Liu and Liang Qin},
  title = {{Improved Algorithm for Insulator and Its Defect Detection Based on YOLOX}},
  booktitle = {Sensors},
  year = {2022},
  doi = {10.3390/s22166186},
}
Figure