Page 98 - MSAM-3-4

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

Table 3. MAPE and RMSE of feature extraction and IoU of the different SR methods
Networks Area Perimeter Circularity Aspect ratio Average MAPE (%) IOU
MAPE (%) RMSE MAPE (%) RMSE MAPE(%) RMSE MAPE(%) RMSE
Bicubic 58.89 342.06 24.35 31.28 16.63 0.128 15.54 0.231 28.85 0.661
EDSR 70.90 396.89 74.81 81.68 42.35 0.318 15.96 0.237 51.00 0.616
SRCNN 12.02 96.91 9.58 19.04 13.04 0.111 8.40 0.167 10.76 0.861
FSRCNN 8.78 127.34 7.15 20.42 9.74 0.090 6.70 0.134 8.09 0.890
VDSR 6.99 84.49 5.97 14.41 8.51 0.081 6.09 0.127 6.89 0.905
RCAN (CA) 10.27 83.56 7.10 16.47 8.81 0.089 5.79 0.123 7.99 0.889
RCAN (CBAM) 5.23 53.12 4.61 12.53 7.64 0.077 4.82 0.121 5.57 0.926
RCAN (SENet) 7.63 78.08 5.22 13.29 7.97 0.081 5.14 0.131 6.49 0.912
U-Net branch 4.79 52.78 4.63 11.63 8.67 0.080 4.98 0.122 5.77 0.925
RCEN branch 3.82 56.94 4.14 12.50 7.19 0.072 4.51 0.122 4.92 0.935
MPSR-Net 3.99 54.76 3.74 11.57 6.11 0.066 4.23 0.112 4.52 0.939
Abbreviations: CA: Channel attention; CBAM: Convolutional block attention module; EDSR: Enhanced deep super-resolution network; FSRCNN: Fast
super-resolution convolutional neural network; IoU: Intersection over union; MAPE: Mean absolute percentage error; MPSR-Net: Melt pool
super-resolution network; RCAN: Residual channel attention network; RCEN: Residual channel with the efficient channel attention network;
RMSE: Root Mean Square Error; SENet: Squeeze-and-excitation networks; SR: Super-resolution; VDSR: Very deep super-resolution network.

4. Conclusion Funding

To overcome the hardware limitations of monitoring This work was supported by the National Natural Science
equipment and noise interference during the L-PBF Foundation of China (Grant No.: 52175481) and the
process, and obtain high-quality melt pool images, this National Natural Science Foundation of China (Grant No.:
paper proposes a dual-path melt pool SR reconstruction 52175528), and in part by the National Key Research and
network. It is composed of RCEN and U-Net branches, Development Program of China, the Chinese Ministry of
considering the characteristics of melt pool images. The Science and Technology (Grant No.:2018YFB1703200).
network achieved high-performance melt pool image SR
reconstruction through attention mechanisms and multi- Conflicts of interest
scale feature fusion, which ensures that high-frequency The authors declare that they have no competing interests.
information in the melt pool image is well captured and the
overall morphology features of the melt pool are effectively Authors’ contributions
extracted. It could also effectively reduce blur and noise in Conceptualization: Kunpeng Zhu, Xin Lin
melt pool images. The PSNR of the network reached 36.13,
and the SSIM reached 0.962. The results showed that the Formal analysis: Xin Lin
Investigation: Kunpeng Zhu, Lei Wu
proposed method not only performed well in traditional Methodology: Yangkun Mao, Lei Wu, Kunpeng Zhu
SR metrics but also achieved good results in melt pool
feature extraction, with the average reconstruction error Writing–original draft: Yangkun Mao, Lei Wu, Kunpeng Zhu
Writing–review & editing: Xin Lin
of melt pool features at 4.52% and the IoU of melt pool
contour reconstruction at 0.939. The network ensures that Ethics approval and consent to participate
after SR reconstruction, the accuracy of melt pool feature
extraction can be improved by upgrading motion blur Not applicable.
and other issues in the LR image. This provides a feasible Consent for publication
method for low-cost monitoring of in situ processes for
software-based L-PBF. Not applicable.

Acknowledgments Availability of data
The authors are grateful to FUH Ying Hsi Jerry of the The raw/processed data required to reproduce these
National University of Singapore for providing valuable findings cannot be shared at the time of publication as the
assistance. data also forms part of an ongoing study.

Volume 3 Issue 4 (2024) 11 doi: 10.36922/msam.5585