Page 95 - MSAM-3-4

Materials Science in Additive Manufacturing Super-resolution method for L-PBF

Pixel upsampling to obtain the final SR image. The structure of four different powers are collected, 70% for training, and
MPSR-Net is shown in Figure 6. LR images will be converted 30% for validation and testing. Figure 7 shows the data
into SR images of ×4 size through the MPSR-Net. distribution of the melt pool features. The bar chart displays
The MPSR-Net can be described by the following the dataset’s histogram while the dashed line represents its
equations: normal distribution fitting. The melt pool features in the
dataset exhibit characteristics of an approximately normal
Out RCENI (IX) distribution, which serves to mitigate both overfitting and
LR
1
Out UnetI (X) underfitting issues during model training. This enhances
the model’s generalization capacity.
2
LR

2
1,
I SR UpSampleConv ConcatOut Out (XI) The training settings of the model are shown in
33

Table 1. In this study, batch normalization and Relu

696
Where I R 12424 is the input LR image; I R 19 activation function were used to enhance the non-linearity
SR
LR
is the output SR image; RCEN (•) and Unet (•) represent of the network and prevent overfitting and gradient
the RCEN branch and U-Net branch, respectively; vanishing. Before training, the data were converted
UpSample (•) represents the Sub-Pixel layer with a to YCbCr channels. Only the Y channel was used and
magnification of 4; and Concat (•) represents the channel normalization was performed on it to ensure the stability
stacking operation. of training convergence and improve model generalization.
Furthermore, to prevent overfitting, considering that the
3. Results and discussion background of the melt pool image was relatively simple,
the maximum training epoch for each model was specified
3.1. Model performance as 50 to ensure that the model will not overfit under the
Based on the experimental platform, a total of 3847 melt premise of convergence of the loss function and testing set
pool images from the single-track melting processes under indicators without fluctuations.

Figure 6. The dual-path melt pool super-resolution network
Abbreviations: LR: Low resolution; SR: Super-resolution

Figure 7. Melt pool features distribution of the dataset

Volume 3 Issue 4 (2024) 8 doi: 10.36922/msam.5585