4.4.2 Detection results under 0.5 annotation percentage
This section provides a summary of the detection results obtained using an APC value of 0.5. Table 4 presents quantitative results, including values for AP, AP@0.75, and AP@0.5. The visualized prediction results are illustrated in Figure 9 (I).
Rows 1-5 of Table 4 present results for one-stage detectors based on PN learning. YOLO v5 achieved an AP value of 62.2%, demonstrating a significant advantage over the other one-stage detectors. In contrast, the two-stage Faster RCNN method outperformed the aforementioned one-stage methods, achieving an AP value of 69.35%, an AP@0.75 value of 80.88%, and an AP@0.5 value of 88.85%.
The PU-based methods in rows 7-9 can be viewed as the combination of Faster RCNN and PU loss. These PU-based methods have achieved better performance than Faster RCNN, as shown in the sixth row. The AP metric of Pi-Index is 3.59% higher than that of Pi-GS and 3.04% higher than that of Pi-FT. A comparison of Tables 2-4 reveals that the performance of each method deteriorated to a certain extent as the labeling ratio decreased. However, the degree of deterioration for Pi-Index is relatively small compared to other mainstream methods. In other words, the performance improvement of Pi-Index is more significant than other mainstream algorithms.
Figure 9 (I) shows the detection results of the seven above methods with an APC of 0.5. As shown in (a) of the first column, the image contains one insulator string and a total of 11 insulators. The insulators at positions (1), (3)-(5), and (7)-(11) are labeled as “Good”, whereas the insulator at position (6) is labeled as “FlashDamaged”.