PSI - Issue 70

Anjireddy Mummadi et al. / Procedia Structural Integrity 70 (2025) 417–423

422

Epoch Vs Intersection over Union(IoU) for Faster RCNN and Mask RCNN

1.2

1

0.8

0.6

0.4

Faster RCNN Mask RCNN

0.2

0

Intersection over Union(IoU)

0

5

10

15

20

25

Epoch

Fig.4. Intersection over Union (IoU) for Mask RCNN and Faster RCNN

Figure 4 highlights the variation in Intersection over Union (IoU) across different epochs. For Faster R CNN, IoU starts at 0.2 in epoch 1 and gradually improves to 0.56 by epoch 25, reflecting steady enhancements in bounding box accuracy. In contrast, Mask R CNN begins with a significantly higher IoU of 0.43 and experiences a sharp increase, reaching 1.15 by epoch 25. This steep rise suggests that Mask R CNN provides more refined object localization compared to Faster R CNN. Standard deviation with different batch sizes (32 and 64), and confidence interval studies with different test -train splits (70 %b-305 ;80 % -20 %) are included in computing accuracy of the Mask RCNN model. The optimum is reported with a validation accuracy of 86% for batch size of 32 and train-test as 80-20. The Mask RCNN model is deployed on the real time bolted joint images and the validation accuracy achieved is 90.3%. The feature map of the first layer is of the CNN is presented in figure 5.

Fig.5. Feature map

7. Conclusions

The key takeaways from the current work can be concluded as; Mask R CNN out performs Faster R CNN in both evaluation metrics i.e mAP and IoU , making it the superior choice for accurate object detection and segmentation. The performance gap widens over training epochs, indicating that Mask R CNN benefits more from extended training. Since Mask R CNN achieves a much higher IoU compared to Faster R CNN, it suggests Mask R CNN excels in fine grained object segmentation, making it particularly useful for applications requiring precise boundary detection. If precision and segmentation are the primary objectives, Mask R CNN is the optimal model, as it consistently delivers better accuracy and localization. The real- world imagery is validated with the Mask RCNN model to classify the bolts as loose or tight.