PSI - Issue 42

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

945

Satyajit Dey et al

3

Fig. 2. Flow chart of model training for image segmentation

Fig. 3. Schematic showing the structure of the UNet architecture

The image segmentation model developed in this work uses a UNet architecture [17]. Figure 3 shows a schematic of UNet. It is comprised of an encoder and a decoder block. The UNet architecture has a ”U-shape” consisting of a contracting path (encoder block) and an expansion path (decoder block). The encoder block consists of multiple pooling layers and layers of encoders which perform convolution operations to reduce the features in the image, while the decoder block performs upsampling of the image until a predicted mask for the image is obtained, which is of the same size as the image. Each encoder is a multi-layer CNN. Residual Net or resnet [18] is a particular type of neural network which employs skip connections increasing accuracy and e ffi ciency of the network. Figure 4 demonstrates the di ff erence between a plain neural network encoder and a resnet. Figure 5 shows a schematic of the convolution process. An input image array (x) of 4x4 is operated on by a 3x3 convolution filter (W). The output array is given by Equation 1. ( W ∗ x ) = y 0 ( i = 0 W i x i ) y 0 ( i = 1 W i x i ) y 0 ( i = 3 W i x i ) y 0 ( i = 4 W i x i ) (1)