In the first column, from (b) to (h), all methods except for M/D-YOLO v4
and YOLO v5 correctly identified the insulator string. However, compared
to Pi-FT and Pi-Index, M/D-YOLO v3 had a larger localization error for
insulator strings. Specifically, the right boundary of the predicted box
by M-YOLO v3 was inaccurate, while the upper boundary of the predicted
box by D-YOLO v3 exceeded the GTBox.
For the relatively small insulators, all methods correctly classified
the insulators at positions (1), (3), and (10). The positions that are
easily undetected or misclassified by these methods are (6), (7), and
(11). According to the IDID’s annotations, M-YOLO v3 missed the
insulators at positions (4), (6), and (7). D-YOLO v3 misclassified the
“FlashDamaged” insulator at position (6) as “Good” and missed the
insulators at positions (7) and (11). M-YOLO v4 missed the insulators at
positions (6) and (7), and further misidentified the “Good” insulator
at position (11) as “FlashDamaged”. D-YOLO v4 missed 3 insulators at
positions (5), (7), and (8). YOLO v5 missed the insulator at position
(9) and had false alarms at positions (8) and (11), in which the
“Good” insulators were identified as “Good” and “FlashDamaged”.
Both Pi-FT and Pi-Index detected the “Good” insulator at position (11)
as both “Good” and “FlashDamaged”.
In summary, Pi-FT and Pi-Index were the most effective algorithms for
labeled data. It is worth noting that the insulator at position (2) was
not labeled in the IDID dataset, yet only Pi-Index was able to identify
it. This demonstrates that Pi-Index is more effective for unlabeled data
than other mainstream algorithms.
In
the second column of Figure 9 (I), there is one insulator string
containing 12 insulators. The insulators at positions (1) and (12) were
only partially visible in the image, while the rest were entirely
visible with clear edges. Specifically, positions (1)-(4), (6), and
(8)-(12) were labeled as “Good”, whereas positions (5) and (7) were
labeled as “Broken”.
The
insulator string was identified only by D-YOLO v4, Pi-FT, and
Pi-Index, whereas the rest of the
methods missed it. Moreover, we focused on the detection details of the
insulator string by D-YOLO v4, Pi-FT, and Pi-Index. D-YOLO v4 predicted
a large error in locating the insulator string, and specifically its
upper boundary exceeded that of the GT-Box. Pi-FT had false alarms on
the prediction of the insulator string. The insulators at positions
(1)-(5) and (3)-(12) were also misclassified as two separate insulator
strings. However, Pi-Index has the best performance on the localization
of the insulator strings. It correctly detected the insulators at
positions (1)-(12) as one insulator string, and the boundary box
basically overlaps with the GT-Box.
For the small-size insulators, all methods correctly detected the
insulators at positions (2)-(4), (6), and (9)-(11). The differences of
detection results from the above methods are mainly concentrated at
positions (1), (5), (7), (8), and (12). Both M-YOLO v3 and D-YOLO v3
missed three insulators at positions (1), (5), and (12), but M-YOLO v3
further misidentified the “Good” insulator at position (7) as
“Broken”. M/D-YOLO v4 missed the insulators at positions (1) and (12).
Pi-FT missed the insulators at positions (1) and (8), while our Pi-Index
only missed the insulator at position (1). YOLO v5 successfully detected
all insulators, performing better than other mainstream methods. In
summary, Pi-Index is significantly better than other mainstream
algorithms except for YOLO v5.
When comparing Pi-Index and YOLO v5, we can be concluded that Pi-Index
is significantly better than YOLO v5 in terms of the large-scale
insulator string. However, Pi-Index has a higher miss rate than YOLOv5
in terms of the small-size insulators. Besides, Pi-Index outperforms
YOLOv5 in terms of locating small targets. This is specifically
reflected in the overlap rate between the predicted box and the GT-Box
for the insulators at positions (5) and (7), where Pi-Index has a higher
rate than that of YOLOv5.