to Table 5, Pi-GS, Pi-FT, and Pi-Index are individually 2.17%, 0.81%,
and 5.25% higher than Faster RCNN’s AP metric, indicating that PU
learning can effectively improve the detection performance under severe
label missing conditions.
Figure 9 (II) displays the
detection results of the seven methods with an APC value of 0.3. The
first scenario, shown in the first column of Figure 9 (II), contains an
insulator string that consists of six insulators. The insulator at
position (6) is partially exposed, while the rest are clearly visible in
the image. The insulators, except for the one labeled as “Broken” in
position (3), are labeled as “Good”.
For the insulator string, M/D-YOLO v3 did not recognize the insulator
string, as shown in (b) and (c). Meanwhile, the other methods were able
to identify it successfully. However, D-YOLO v4 and YOLO v5 had large
localization errors when detecting the insulator string. Specifically,
the left and right boundaries of the predicted box by D-YOLO v4 both
exceeded the boundaries of the GTBox. The predicted box by YOLO v5 was
completely enclosed within the GTBox. For small-size insulators, M-YOLO
v3 and M-YOLO v4 missed all insulators. D-YOLO v3 and D-YOLO v4 detected
only two insulators, while YOLO
v5 and Pi-FT detected three insulators. Therefore, Pi-Index outperformed
the other methods with four insulators detected correctly.
There is one insulator string and seven insulators in the second
scenario, shown in the second column of Figure 9 (II). The insulators
are arranged closely and the positions (1)-(5) labeled as “Good”,
while the insulator at position (7) labeled as “FlashDamaged”. From
the detection results of the insulator string, it can be seen that the
results from all the methods are poor. Only Pi-FT and Pi-Index were able
to successfully recognize the insulator string. For the insulators,
M/D-YOLO v3 and M-YOLO v4 did not detect any insulators, while D-YOLO v4
only detected the insulator at position (3). This indicates that these
methods have a high rate of missed detections. The performance of YOLO
v5 and Pi-FT is slightly better. In detail, they only missed the
insulator at position (7). Pi-Index further successfully identified the
insulator at position (7). The detection results demonstrated that
Pi-Index performed best for detecting insulators in contrast to the
other detectors.