Page 97 - MSAM-3-4

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

Root Mean Square Error (RMSE) values of the melt pool the melt pool greatly interfere with the feature extraction.
features between the SR and HR images were calculated Simple interpolation methods cannot improve the low
as evaluation metrics. Melt pool features include the melt quality of melt pool images resulting from motion
pool area, perimeter, circularity, aspect ratio, and the blur and other factors. The proposed method achieved
average MAPE of these metrics. Meanwhile, the melt pool the minimum average MAPE and a high IoU score of
contour IoU of the HR and LR images was also calculated, 0.939 in the overall evaluation, and the comprehensive
as shown in Figure 9. performance of MAPE and RMSE of MPSR-Net was the
best, demonstrating its accuracy of the feature extraction
As shown in Table 3, SR based on DL can improve the for the melt pool. This can provide a solid foundation for
accuracy of melt pool feature extraction. The blurring and the subsequent monitoring, identification, and control of
ghosting effects caused by the dynamic characteristic of the metal AM process.

Figure 8. The SR results comparison between different SR methods
Abbreviations: CA: Channel attention; CBAM: Convolutional block attention module; FSRCNN: Fast super-resolution convolutional neural network; HR:
High resolution; LR: Low resolution; MPSR-Net: Melt pool super-resolution network; RCAN: Residual channel attention network; SR: Super-resolution;
VDSR: Very deep super-resolution network

Figure 9. Feature extraction of the melt pool
Abbreviations: h: The height of the melt pool; HR: High resolution; IoU: Intersection over Union; OTSU: Nobuyuki Otsu’method; SR: Super-resolution;
w: The width of the melt pool

Volume 3 Issue 4 (2024) 10 doi: 10.36922/msam.5585