The detection results of (b)-(h) show that all methods have identified
the insulator string. However, there was a false alarm in M-YOLO v3, and
three more insulator strings were detected in detail. Meanwhile, the
predicted boxes from D-YOLO v3, M-YOLO v4, and D-YOLO v4 exhibit a
relatively larger error on the right border. On the contrary, YOLO v5,
Pi-FT, and Pi-index provide the more accurate localization for the
insulator string.
In the detection of insulators, YOLO v5, Pi-FT and our proposed Pi-Index
correctly identified seven insulators (annotated in IDID) and the other
methods had missed insulators to varying extent. This demonstrates that
the Pi-Index achieved competitive results on the annotated insulators in
IDID. Furthermore, the IDID dataset lacks annotations at positions (6),
(7), (10), (11), and (12). Based on the aforementioned analysis, we
further compared the performance of YOLO v5, Faster RCNN, and Pi-Index
on unannotated insulators. recognized the unannotated insulators at
positions (7) and (10). Therefore, both annotated and unannotated
experimental results proved the effectiveness of Pi-Index.
4.4.2 Detection results under 0.7 annotation
percentage
A
decrease in data annotations inevitably results in a decline in the
performance of the model. This circumstance facilitates a more
comprehensive exploring the performance of different algorithms under
imperfect labeling conditions. In this section, we further increase the
proportion of unlabeled targets in the dataset. In detail, we randomly
delete 30% of the labels in the IDID dataset. The detection results of
different methods are organized in Table 3, and the visualization of
annotated and predicted boxes is shown in Figure 8 (II).
From Tables 2 and 3 (1.0 vs. 0.7 APCs), it can be seen that each
method’s performance in Table 3 has declined compared with the
counterpart in Table 2. Meanwhile, the performance differences between
the various methods become more significant while maintaining
consistency in trends. As detailed in Table 3, YOLO v5 surpassed other
one-stage detectors (rows 1-4) with an AP value of 72.2%. Nevertheless,
the AP, AP@0.75, and AP@0.5 of Faster RCNN are 3.84%, 1.08%, and
1.66% higher than those of YOLO v5, respectively.
Compared with Faster RCNN, the AP
value of Pi-GS, Pi-FT, and Pi-Index have shown improvements of 2.56%,
2.91%, and 3.53%, respectively. This indicates that the addition of PU
loss can effectively alleviate the effect of missing labels. From the
last three rows in Table 3, Pi-Index achieved the best performance with
an AP value of 79.57%, an AP@0.75 of 89.45%, and an AP@0.5 of 92.64%,
which also outperformed other mainstream algorithms in each category’s
AP. As depicted in Table 2, when compared to Pi-GS and Pi-FT, Pi-Index
exhibits a marginal improvement in the AP metric by less than 0.50%.
But in Table 3, the AP of Pi-Index increased by 2.56% and 2.91%
compared to those of Pi-GS and Pi-FT, respectively. This proves that the
Pi-Index can achieve more significant performance improvements as the
amount of unlabeled data increases.
Figure
8 (II) displays the detection results of seven methods, which are
identical to those depicted in Figure 8 (I), under an APC of 0.7.
The two columns in Figure 8 (II)
correspond to two images with insulators. The first image or column
contains a total of two insulator strings and 12 insulators. The
insulators within the left-hand string are distributed evenly and appear
relatively large in the image. Whereas the insulators in the right
insulator string are densely arranged, and some insulators therein are
occluded due to shooting reasons. Consequently, distinguishing the
insulators located on the right side poses a greater challenge compared
to those on the left side. According to the IDID dataset, the insulators
at positions (2), (3), (5), (6), and (7)-(12) have been annotated as
“Good”, while the insulator at position (4) has been labeled as
“Broken”.
From (b)-(h) in the first column of
Figure 8 (II), all methods have successfully identified the insulator
string on the left, whereas they failed to identify the insulator string
on the right. This discrepancy is presumably due to a greater degree of
occlusion affecting the insulator string on the right. Both M/D-YOLO v3
and M/D-YOLO v4 exhibited larger localization errors for the left
insulator string compared to other methods. Specifically, the
lower bounds of the predicted boxes
for M/D-YOLO v3 and D-YOLO v4 exceeded the ground-truth lower bound by a
large amount. The localization error of M-YOLO v4 was reflected in the
larger predicted box, which enclosed the GT-Box. Furthermore, the
Pi-Index possessed the highest overlap between its predicted box and the
GT-Box than YOLO v5 and Pi-FT. These results indicate that the Pi-Index
is more accurate for localizing insulator strings.
For small target insulators, M-YOLO
v3 and YOLO v5 exhibited subpar performance with an accuracy rate below
50%. D-YOLO v3, M-YOLO v4, and Pi-FT successfully detected seven
insulators. The first six insulators detected by all of them are
positions (3)-(5), (7), (9), and (10). The seventh insulator detected by
D-YOLO v3, M-YOLO v4, and Pi-FT is position (1), (2), and (12),
respectively.
Furthermore, D-YOLO v4 and Pi-Index
have correctly identified 8 insulators, indicating that these two
methods outperformed the previously mentioned methods. For the
localization of the left insulator strings, the predicted box of D-YOLO
v4 was shifted to the right side of the GT-Box. Additionally, the lower
boundary of the predicted box extended beyond the GT-Box. Contrastingly,
Pi-Index generated a predicted bounding box that exhibited a large
overlap with the GT-Box. Furthermore, we conducted a comparative
analysis between D-YOLO v4 and Pi-Index based on their respective
performance in detecting insulators within insulator strings. In the
case of the left insulator string, both D-YOLO v4 and Pi-Index correctly
identified the insulators at positions (1)-(4). However, D-YOLO v4
exhibited a large offset when locating insulators at positions (1) and
(2). Within the right insulator string, both D-YOLO v4 and Pi-Index
detected four insulators. Despite these detected insulators being
located at different positions, there was a significantly greater
overlap between GT-Boxes and the predicted bounding boxes produced by
Pi-Index than those produced by D-YOLO v4. In conclusion, the proposed
Pi-Index exhibits superior performance to the comparison algorithm in
terms of both the number of correct detections and localization
accuracy.
The
second column of both Figure 8 (I) and (II) shared the same image or
detection scenario. This scenario
consisted of one insulator string and 12 insulators. It is worth noting
that the same scenario was applied with different data APCs during the
training process, which led to different detection results. The
following analysis is based on APC = 0.7.
In Figure 8 (II), with the
exception of M-YOLO v3, all other methods failed to successfully
identify the insulator string. This may be attributed to the fact that
during the process of reducing the APC, labels for insulator strings
within similar scenes were randomly eliminated, resulting in their being
interpreted by the network as background. However, M-YOLO v3
misclassified the insulators at positions (2)-(7) as an insulator
string. Combining the detection results of M-YOLO v3 under APC=1 (in the
second column of Figure 8 (I)), it can be seen that M-YOLO v3 identified
up to four insulator strings. These identified insulator strings only
encompassed a portion of the insulators present within the annotated
insulator string. Therefore, we may conclude that M-YOLO v3 has a
tendency to identify a series of consecutive insulators as forming an
insulator string, rather than detecting the entire insulator string.
Contrastingly, the remaining methods failed to detect the insulator
string.
Table 4: Detection results
of our method and the baseline method with (APC=0.5) in the training
process.