4 Conclusion
In this paper, we propose a framework to explore insulator defect
detection under circumstances that combine incomplete annotation and
sample imbalance. The framework introduces a PU-RPN that integrates
improved PU learning with RPN and incorporates focal loss into the ROI
Head. On one hand, the improved PU loss is used to address the problem
of incomplete annotation by appropriately calibrating the losses of
different samples. In addition, the proposed Pi-Index strategy is
responsible for estimating a more accurate class prior by combining
classification confidence scores from RPN and predicted boxes from
ROIHead. On the other hand, focal loss is incorporated into the ROI Head
to alleviate performance degradation caused by sample imbalance. To
verify the effectiveness of our proposed framework, we conducted two
groups of experiments. The experimental results demonstrate that our
method outperforms not only the baseline method or Faster RCNN, but also
other mainstream methods. Specifically, our method achieved the highest
AP metrics (88.11% for 1.0 APC, 79.57% for 0.7 APC, 73.49% for 0.5
APC, and 63.82% for 0.3 APC) with different proportions of annotations
when compared to mainstream methods.
Funding Statement: This research is partially supported by the
R&D Program of Beijing Municipal Education Commission (KM202210009003),
the National Natural Science Foundation of China (62001009), and the
Scientific Research Foundation of North China University of Technology.
Author Contributions: Author Contributions: Conceptualization,
F. Q. P.; Methodology, F. Q. P., C.Y. L. and J. S. Z.; Validation, J. S.
Z., C. Y. L.; Writing—original draft preparation, F. Q. P., and C. Y.
L; Writing—review and editing, J. S. Z., and C. Y. L. All authors have
read and agreed to the published version of the manuscript. All authors
have read and agreed to the published version of the manuscript.
Availability of Data and Materials: The Insulator Defect Image
Dataset refers to:
https://ieee-dataport.org/competitions/insulator-defect-detection
Conflicts of Interest: The authors declare that they have no
conflicts of interest to report regarding the present study.