1 Introduction
Transmission of electricity throughout the power grid is accomplished
with the assistance of high-voltage transmission lines, electrical
insulators, and power towers. Although the insulators therein are not
directly responsible for delivering electricity, they suspend the
overhead transmission lines and prevent the transmission lines’
grounding [1]–[5]. Furthermore, they must constantly withstand
the supply voltage and bear the load of the transmission lines’ gravity.
Therefore, the insulator is an essential component in a high-voltage
power system, which facilitates maintaining the safety and stabilization
of the power grid.
Because of the extensive deployment of the power grid, the insulators
need to be adaptable to different kinds of natural environments and
geographic conditions. A fairly large number of them are exposed to an
outdoor environment throughout the entire year, which is more
susceptible to the erosion caused by harsh climates such as sunlight,
rain, and snow [6]. Besides, overvoltage shocks from lightning and
on-off operations, mechanical load, weight of wires, as well as metal
accessories may also make the insulators more easily prone to self-blast
or breakage [7]. The above factors inevitably give rise to the
insulators’ defects and reduce their service lifetimes. Moreover, aged
and defective insulators may cause regional grid failures and enormous
economic losses if they are not periodically inspected [2].
One solution is manual inspection of power equipment, but this
traditional method is considered low-efficiency, labor-intensive, and
unsafe. It is incapable of providing quick feedback on the condition of
insulators and meeting the inspection needs of a modern smart grid,
namely more frequent inspection, repair, and maintenance [8]. Due to
the low cost, miniaturization, and high mobility of Unmanned Aerial
Vehicle (UAV) inspection, it has replaced manual inspection and become
the mainstream inspection scheme for power equipment [9]–[12].
For insulator defect detection, the UAV inspection collects a large
number of aerial images with insulators. The collected insulator images
not only enable the automatic identification of the insulator defects,
but also make it an urgent problem to be solved.
Early algorithms for insulator defect detection are based on handcrafted
features and machine learning technologies [1], [13], [14].
However, feature design is time-consuming and costly, and requires the
assistance of experienced experts. In recent years, deep learning
technology has made a breakthrough in image classification [6],
object detection [15], image segmentation [16], etc. The deep
learning-based detectors were introduced in this field
[17]–[20], including both You Only Look Once (YOLO) family
[21]–[24] and Region-based CNN (RCNN) series [25]–[27]
detectors. On the one hand, the YOLO-based detectors were investigated
to enhance the speed of insulator detection by incorporating a
lightweight backbone, such as, replacing the backbone by MobileNet
[28] in YOLO v3 [23] and YOLO v4 [24], and choosing suitable
backbone among the four versions of YOLO v5 [9]. Simultaneously,
attention mechanisms have been introduced into insulator detection
research. In detail, channel-wise self-attention was merged with
TinyYOLO v4 to facilitate feature representation [10], while YOLO v5
pipeline was incorporated with a triplet attention module in [9] and
a Convolutional Block Attention Module (CBAM) in [29] for providing
more context information for insulator defect detection.
On the other hand, Faster RCNN, a two-stage detector, was introduced
into the insulator detection community [17], [30]. It first
roughly generates the proposals of insulators and then refines the
proposals for locating the insulators’ defects. Moreover, Zhong et al.
modified the standard Faster RCNN pipeline to consider arbitrarily
oriented insulator localization [31]. In [32], the attention
mechanism was introduced in Faster RCNN for self-explosion insulator
defects. In summary, the aforementioned approaches focused on the
following problems: 1) A complex background and various types of
insulators make insulator defect detection difficult; 2) some severely
damaged insulators are too small to recognize; 3) detection speed and
accuracy needs to be balanced.
In most fields of object detection, ”perfect annotation” indicates that
the labeled bounding boxes are close to the targets’ true boundaries and
there are no missing labels. Therefore, the acquisition of the perfect
annotation is time-consuming and labor-intensive. To decrease the
dependency on perfect annotation, numerous research works have studied
the different scenarios of imperfect annotation, such as incomplete
annotation [33], [34], unreliable labels [35], and
incremental new categories [36]. However, the incomplete annotation
problem of insulator defect detection is not thoroughly investigated.
Besides, there is a sample imbalance among different categories in our
task.
In this paper, we propose a novel framework for incomplete annotation
and sample imbalance in insulator defect detection. The proposed
framework is based on the Faster RCNN detector and integrates Positive
Unlabeled (PU) learning and focal loss. It is termed Pi-Index that
follows the name of the proposed algorithm for estimating the class
prior, a key parameter in PU loss, to improve PU learning. The algorithm
is designed to generate a continuous value (Pi-Index) for each anchor as
its probability of being positive. In the aspect of network
architecture, Region Proposal Network (RPN) is improved by introducing
PU learning to overcome the problem of incomplete annotation, whereas
the focal loss strategy is applied to Region Of Interest (ROI) Head, to
alleviate the impairment caused by sample imbalance.
The contributions of this paper are summarized in the following aspects:
1) A novel estimation strategy for the class prior, termed Pi-Index, is
proposed to improve vanilla PU learning; 2) The PU learning strategy and
focal loss are separately incorporated with RPN and ROI Head, which are
responsible for the above two problems; and 3) Experiment results show
that the proposed framework achieves better performance compared with
baseline methods when missing labels or sample imbalance scenarios
occur.
The rest of this paper is organized as follows. Section 2 describes the
proposed framework that contains FPN backbone, RPN with PU learning and
ROI Head with focal loss. Section 3 is about experimental results that
prove the effectiveness of the proposed framework. Section 4 gives a
summary of this paper.