PSI - Issue 44

Agnese Natali et al. / Structural Integrity Procedia 00 (2022) 000–000 7

Agnese Natali et al. / Procedia Structural Integrity 44 (2023) 2020–2027

2026

Fig. 9. Patch extraction on PAVIS-FABRE dataset for classification. (a) Sliding window-based patch extraction, (b) distribution of the extracted patches for different classes.

We observe that detections are meaningful and have significant overlap with the annotated regions. Nevertheless, the CODEBRIM and PAVIS-FABRE datasets have different classes. Moreover, not all the classes in PAVIS FABRE dataset correspond to defects, therefore, we have to train another classifier suitable for this dataset. To do so, we follow a patch-based training approach by extracting 227x227 sized patches at a stride of 64 pixels from all the annotated area. In addition, we also extract patches from the unannotated regions and assign them to the “Background” class. Distribution of the extracted patches for different classes is shown in Fig. 10 (b).

Fig. 10. Adapting our model to the PAVIS-FABRE dataset. We freeze all the layers, except the last fully connected layer (fc2) of our trained model. We then replace fc2 with a different classification head, according to the classes in the PAVIS-FABRE dataset and train it using the patches extracted from this dataset. The idea is to then replace the classification head, which is the last layer in our trained model, with a different classification head needed for the PAVIS-FABRE dataset. The extracted patches can then be used to update our model by training only the replaced classification head as depicted in Fig. 11. The training is performed using the same setup as discussed in Section 3.1. The qualitative results obtained with the updated model are shown in Fig. 12. The results shown in Fig. 12 are encouraging as various infrastructure elements are detected reasonably. The results may be improved by using various data augmentation techniques during training. Moreover, the statistics presented in Fig. 7 and Fig. 10(b) can help in identifying less annotated classes. Accordingly, inclusion of more annotated data may lead to further improvement in the results.