Rows
1-6 of Table 2 display the detection results of PN-based, primarily from
one-stage (rows 1-5) and two-stage (row 6) networks. YOLO v5
significantly outperforms other one-stage networks with an Average
Precision (AP) of 84.50%, an AP@0.75 of 96.20%, and an AP@0.5 of
98.30%. However, the two-stage Faster RCNN achieved an AP of 87.19%,
which is approximately 2.69% higher than YOLO v5. This demonstrates
that Faster RCNN exhibits the best performance among the above PN
learning frameworks.
PU-based methods with Faster RCNN as
the backbone are represented in rows 7-9 of Table 2. They have improved
detection performance in comparison with PN-based detectors. Among these
methods, our proposed Pi-Index obtained the highest performance with an
AP metric of 88.25%, which to some extent indicates the advantage of
Pi-Index in estimating the class prior. Moreover, the AP values of Pi-GS
[49] and Pi-FT [34] are 0.41% and 0.18% lower than Pi-Index,
respectively. The reason may be that their class prior estimates are
based on a hyper-parameter or a fixed threshold. This parameter or
threshold is determined via grid search, but the search interval of 0.1
may stride over the optimal value and further lead to performance
deterioration.
Figure 8 (I) shows the detection results of various methods with APC=1.
The different columns correspond to different insulator detection
scenarios, while each row is the detection result of a particular
method. Rows (a)-(h) display manual labels and prediction results of
M-YOLO v3, D-YOLO v3, M-YOLO v4, D-YOLO v4, YOLOv5, Pi-FT, and our Pi
-Index. The green solid and red dashed boxes in the figure represent the
Ground-Truth
bounding Boxes (GT-Boxes) and
predicted boxes, respectively. The text in the upper left corner of each
GT-Box and predicted box individually indicates the annotated and
predicted category, including: “String”, “Good”, “Broken”, and
“FlashDamged”. To facilitate the analysis of prediction results,
yellow arrows are employed to number those small and densely arranged
insulators. The analysis of detection results is divided into two
aspects: 1) The count of correctly detected insulators, i.e., whether
any insulator strings or insulators were missed or detected repeatedly;
2) The localization accuracy of predicted boxes for the insulator
strings or insulators.
In
the first column of row (a), there is one insulator string and 13
insulators. In detail, the insulators at positions (1)-(4) and (9)-(12)
are labeled as “Good”, while those at positions (5), (7), and (13) are
labeled as “Broken”. However, the insulators at positions (6) and (8)
are not labeled. All methods
successfully identified the insulator string. Furthermore, the
comparison of (b)-(e) and (f)-(h) reveals that M/D-YOLO v3 and v4
exhibit larger errors than YOLO v5 and two-stage algorithms in locating
the string.
For the small-scale insulators, all
methods correctly identify and accurately locate those at positions
(1)-(3) and (7). The detection errors of each method are mainly
concentrated on positions (5) and (13). Primarily, the “Broken”
insulator at position (5) is misidentified as “Good” by M-YOLO v3,
D-YOLO v3, and YOLO v5, misclassed as “FlashDamaged” by D-YOLO v4, and
repeatedly identified as both “Good” and “Broken” by Pi-FT. Only
M-YOLO v4 and the method proposed in this paper manage to correctly
identify it. Next, the “Broken” insulator at position (13) was missed
by M-YOLO v3, M-YOLO v4, and YOLO v5. It was misidentified as “Good”
by D-YOLO v3 and “FlashDamaged” by D-YOLO v4. Only Pi-FT and our
proposed method correctly identified it. By comprehensively considering
the detection results of position (5) and (13), it reveals that M-YOLO
v4, Pi-FT, and Pi-Index demonstrate
superior performance.
Compared with Pi-Index, M-YOLO v4
not only misclassed the “Good” at position (9) as “Broken” but also
missed the insulators at positions (10)-(12). In the detection results
of Pi-FT, false alarms occurred at positions (4) and (5). They
individually belong to the categories of “Broken” and “Good”, but
are simultaneously identified as both “Good” and “Broken”. This
further substantiates the superiority of the Pi-Index over the other two
algorithms. There are another two unlabeled targets at positions (6) and
(8). Neither M-YOLOv4 nor YOLO v5 recognized them, while M-YOLO v3 and
D-YOLOv4 only recognized positions (6) and (8), respectively. By
contrast, D-YOLO v3, Pi-FT, and ours recognized the insulators at
positions (6) and (8). In summary,
our proposed method not only
achieves the overall best
performance for the labels provided by IDID
but also yields competitive results
on unlabeled data.
The second column in Figure 8 (I)
contains one insulator string that is composed of 12 insulators.
However, seven insulators therein were annotated in IDID as either
“Good” or “Broken”. The former category includes six insulators at
positions (1)-(4), (8), and (9), while the latter category includes only
one insulator at position (5). It
is worth noting that the insulator at position (9) was repeatedly
annotated as “Good”.
Table
3: Detection results of our method and the baseline method with
(APC=0.7) in the training process.