Page 92 - MSAM-3-4

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

A B C

Figure 2. Typical low-quality melt pool images: (A) ghost, (B) overexposure, and (C) noise

Figure 3. Degradation of the melt pool image
Abbreviation: LR: Low resolution

lightweight structure and an improved attention module are also beneficial in mitigating gradient vanishing and
to adapt to the simplicity of the melt pool image and the enhancing feature propagation. Based on the above, the
40
importance of its boundary information. It can prevent RCEN consists of N residual group (RG) with a long skip
parameter redundancy and overfitting, and improve model connection. It ensures that low-frequency information
speed. in the melt pool image could be preserved, while high-
frequency information could be mapped to HR space
The RIR structure is based on a residual mechanism,
and the network contains dense links of multiple residual through the RIR structure, thus outputting as melt pool
features. Each RG block consists of M residual channel with
structures. It ensures that while messages are passed the ECA-Net block (RCEB), while RCEB consists of two
between layers, the network can also obtain lower- convolutional (Conv) layers, a Relu function, an ECA-Net,
level raw features as additional input and make this and a residual connection. Based on the characteristics of
information preservation explicit through additive identity melt pool images, N is set to 5, M is set to 10, resulting in a
transformations. Low-level features (especially spatial lightweight model. Meanwhile, the kernel size of the Conv
information) can therefore be integrated with high-level layers in Figure 4 is 3.
features to enhance the reconstruction ability of networks.
32
High-frequency information such as melt pool boundaries According to RCAN, the RCEN branch can be
is mainly reflected in the brightness differences in the described by the following equations:
spatial dimension, so preserving spatial features is crucial. R Conv (II)
I

The RIR structure can ensure that low-level features can be long 33 LR
effectively integrated with high-level features in a multi- R
level jump connection. Meanwhile, skip connections Output Conv 3 3 RG long R long (III)
N
1

Volume 3 Issue 4 (2024) 5 doi: 10.36922/msam.5585