PSI - Issue 42

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

947

Satyajit Dey et al / Structural Integrity Procedia 00 (2019) 000–000

5

Fig. 6. a) Metallographically polished 316L steel SLM and DED samples b) Micrographs of the di ff erent types of defects c) Masked micrographs of the defects

, Fig. 7. Example of (a) image (b) mask

type of defect was masked with a unique colour: chartreuse for lack of fusion, green for porosity and red for micro crack. The background was masked in black. For the study, two sets of image resolution were used: 320x320 and 960x960. For the training and validation data set, images were resized. Horizontal and vertical flip were applied to the training set with a probability of 0.5.

3.3. Model Development

A UNet architecture was constructed with 4 sets of encoder and decoder units. Many types of encoders are available out of which vgg11, senet154 and resnet18 were used for comparison. vgg11 has 9 million parameters, senet154 has 113 million parameters and resnet18 has 11 million parameters.

3.4. Training and Validation

Table 1. Matrix of cases Case

Encoder

Image Resolution

Run Time

1 2 3

resnet18 senet154 resnet18

320x320 320x320 960x960

1h 9m 6h 20s 7h 11m

Model training was performed with three sets of model and input parameters as illustrated in Table 1. The maximum number of iteration was set at 1000 and an optimum batch size of 5 was used. The model parameters corresponding to the lowest validation loss were stored. Training times increase exponentially with the image resolution and also with the number of parameters in the encoder. Table illustrates a comparison of run times for the three cases used.