Page 91 - MSAM-3-4

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

Figure 1. Experimental platform
Abbreviation: ROI: Region of interest

(2) High temperature and high brightness. The melt pool For the melt pool, brightness information is the most
is a high-temperature liquid metal region, typically important. Therefore, the original image was transformed
appearing as a bright area in the image with its from the RGB color space to the YCbCr color space. The
brightness being related to the radiation properties of Y channel (luminance channel) is considered as network
the melt pool. input. The Cb and Cr channels were upsampled using
(3) Texture and edge features. Melt pool images typically bicubic interpolation, then they were combined with the
exhibit unique textures and edge features. These reconstructed Y channel to visualize the final reconstructed
reflect the flow phenomena, heat distribution, and melt pool image. The input and output of the network are
metallurgical reaction processes inside the melt pool. referred to as LR and SR image to simplify the description.
By analyzing these features, a deeper understanding of
the physical mechanisms in the L-PBF process can be 2.3. The residual channel with the efficient channel
gained. attention network (RCEN)
Typical low-quality melt pool images are shown in To reconstruct the geometry of the melt pool, its boundary
Figure 2. information in LR images needs to be obtained. The
boundary of the melt pool is the area with the most drastic
Studies in SR field typically use image degradation
methods to generate LR datasets for training. Considering grayscale changes in the image, representing the high-
38
the dynamic behavior of the melt pool and monitoring frequency information of the image. How to distinguish
noise, the melt pool image degradation methods include: the ghosting or blurring from the LR boundary requires
the downsampling (×4), motion blur, and salt-and-pepper the network to extract high-frequency features more
32
noise. Motion blur is obtained through a motion blur efficiently. Inspired by the RCAN, this study constructed
convolution kernel, and downsampling uses the bilinear a chain structure branch for extracting melt pool image
interpolation method. As shown in Equation I and Figure 3, features, which is called the RCEN, as shown in Figure 4.
I LR I HR k (I) The network was ensured to have enough depth to
4
blur
reconstruct high-frequency information through the RIR
696
where, I HR R 19 is the high-quality melt pool structure. Then, the efficient channel attention network

39
image (grayscale), I R 12424 is the degraded image, k (ECA-Net) was adopted to achieve weight redistribution
blur
LR
represents the motion blur convolution kernel, ↓ between high-frequency and low-frequency information to
4
represents ×4 downsampling, and η represents salt-and- allow for adjustment of the proportion of high-frequency
pepper noise. information. Compared to RCAN, it adopts a more
Volume 3 Issue 4 (2024) 4 doi: 10.36922/msam.5585