PSI - Issue 42

Satyajit Dey et al. / Procedia Structural Integrity 42 (2022) 943–951 Satyajit Dey et al / Structural Integrity Procedia 00 (2019) 000–000

948

6

Fig. 8. Example performance of the di ff erent models

Fig. 9. Bar chart of prediction versus ground truth for di ff erent defects for case 3

4. Results

The trained model was used to detect defects in a set of 50 images. Figure 8 shows a comparison of the predicted masks for all the cases with the ground truth for an image. In general, the prediction is good for all the models with the exception of micro-cracks and only case 3, which uses a high resolution image (960x960), could predict micro cracks. Figure 9 shows a comparison of the total number of predicted versus actual defects in all the 50 images. There is a visible over prediction in all the cases except for pores / porosity. Figure 10 shows the confusion matrix for the model (case 3). All the LoF defects in the actual data were correctly identified by the model. However, there is a small number of over-prediction of LoF defects arising out of misclassifying some porosity defects as LoF. For porosity defects, there is both a small number of under-prediction and over-prediction. Out of a total of 106 micro-cracks in the actual images, 91 were identified correctly. However, a significantly high number of over-predictions were made for micro-cracks. Apart from the number of defects, it is also important to measure if the sizes and shapes of the defects were correctly identified too. Figure 11 shows an example where it can be observed that even if a defect is located correctly, the size or shape of the defect in the predicted mask significantly di ff ers from the ground truth for the image. In order