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

946

4

Fig. 4. Schematic of a plain CNN and a resnet unit

Fig. 5. Schematic illustration of image convolution

3. Model Training

Training of the image segmentation model involves the following steps:

• image acquisition • image processing • model development • training and validation

3.1. Image Acquisition

Selective Laser Melting (SLM) AM process was used to manufacture a square block of 316L samples. A pho tograph of the sample is shown in figure 1. The sample was sectioned in half and the cut surfaces were prepared for microscopy analysis using standard metallographic methods. A Scanning Electron Microscope (SEM) was used to systematically collect high resolution images of the prepared surfaces along the build direction of the samples. A 1500x magnification was used to collect images of 2560*1920 pixels. Figure 6 shows the polished samples, the micrographs of the defects and their corresponding masks.

3.2. Image Processing

A total of 40 images were used for creating the training and validation data sets with a 3:1 split. Masks or ground truth images were manually created for all the images. Figure 7 shows an image with its corresponding mask. Each