PSI - Issue 37

Kafayat Eniola Hazzan et al. / Procedia Structural Integrity 37 (2022) 274–281 Hazzan and Pacella/ Structural Integrity Procedia 00 (2019) 000 – 000

279

6

at least one false positive and 33 % results with at least one false negative. In cases were false positives were high, the IM included defects such as porosity and balling. Future improvement of this method will include reducing these failure modes of the IM and using other techniques to segregate other defect identification.

Fig 5. a) Visual form of intersection calculation for accuracy calculation compared to input SEM. b) Red areas – False negative. Blue area – Broken cracks. Yellow areas – False positive.

IM using simple image processing techniques has many benefits over transfer learning and CNNs; this includes faster processing times and reduced computational demand. The amount of information that can be extracted from the detected regions in CNN based methods are harder to extract and program (Kim et al., 2019). Techniques discussed in this paper are easy to replicate for quick analysis of SEM images with basic computer programming knowledge and limited CPU and GPU capability. 4. Conclusion Image processing algorithms were used to identify and segment crack regions from SEM images of laser engineered WC. The method is able to find cracks, location, and geometry. The results show successful segmentation of cracks from SEM images with an identification accuracy greater than 95 % across a range of different laser processing parameters. Further work will investigate maximising the accuracy and include the identification of other induced defects such as splatter and balling. Acknowledgements The authors would like to acknowledge the Manufacturing Technology Centre for their financial support to the Ph.D. project, and the assistance of Shaun Fowler in the Loughborough Materials Characterisation Centre. Appendix A. Accuracy of Crack identification method across experiments Crack identification accuracy results from laser engineered WC SEM across 49 experiments. Experiment Fluence (J/cm 2 ) Frequency (kHz) No. of cracks Method accuracy (%) 1 0.030 100.0 13 97.2398 2 0.030 17.5 22 97.1326 3 0.030 28.0 24 97.1430 4 0.030 40.0 36 97.1691 5 0.030 5.0 9 96.5749 6 0.030 52.5 21 97.2920 7 0.030 75.0 14 97.4577 8 0.050 100.0 113 96.6740 9 0.050 17.5 101 96.5505 10 0.050 28.0 135 95.7561 11 0.050 40.0 189 96.0298