Page 96 - MSAM-3-4

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

The SR results were evaluated with the PSNR and of the melt pool is enhanced. Further, multi-scale features
SSIM. To verify the superiority of the MPSR-Net in the SR of the melt pool image are extracted by the U-Net branch.
reconstruction of melt pool images, the traditional Bicubic Other attention mechanisms (CBAM, CA, and SENet)
method, SRCNN, FSRCNN, and other DL models were were used for comparison, proving that the advanced
used for comparison. The results are shown in Table 2.
attention mechanisms can effectively improve the
In addition to EDSR, the performance of other DL performance of the model. In addition, ECA-Net is proven
models was better than the traditional interpolation to be the most effective. Compared to only using a single
method, demonstrating a high value of DL in melt pool branch, different key information of the melt pool can be
image reconstruction. The MPSR-Net achieved the best extracted through the dual-path form, and then feature
results compared to other methods, indicating that directly fusion can improve the network performance. In the end,
applying traditional SR models to melt pool images cannot the average PSNR of the model reached 36.13, and the
achieve excellent results. The targeted improvements average SSIM reached 0.962. The SR reconstruction effect
are needed based on the characteristics of the melt pool. was found to be good, which can be used for melt pool
Therefore, this study proposes targeted optimization. It monitoring.
first lightweights the network to ensure no parameter
redundancy or overfitting according to the melt pool Figure 8 shows the comparison of different SR
image with a simple background and target. Meanwhile, methods. It can be observed that the MPSP-Net has clearer
by adopting the attention mechanism to emphasize high- edges of the melt pool, and the reconstructed boundaries
frequency information, the boundary reconstruction effect are also clearer for the case of spatter adhesion. The
reconstructed melt pool image by MPSR-Net restored
more morphological details and effectively overcame the
Table 1. Parameters used in the experiment
interference of noise. Furthermore, it can be difficult to
Parameter Experiment setup assess image quality with subjective visual evaluation. The
Data set 3847 images following section showcases the use of melt pool features
Gradient update algorithm Adaptive moment estimation (Adam) as new evaluation metrics.
Loss function Mean Squared Error (MSE) Loss 3.2. Melt pool features extraction
Learning rate 1e-4
Iterations 100 To demonstrate the significance of SR for melt pool
images, melt pool features were extracted through OTSU
Software environment Python 3.9; PyTorch 2.0.0 threshold segmentation and contour extraction methods. 23
Hardware environment GPU RTX 4070; CPU Intel i7-13700K The HR images were used as ground truth. The MAPE and

Table 2. PSNR and SSIM of different SR methods
Networks 90w 110w 130w 170w
PSNR SSIM PSNR SSIM PSNR SSIM PSNR SSIM
Bicubic 21.91 0.8267 23.36 0.8579 23.01 0.8466 23.23 0.8432
EDSR 31 21.03 0.6170 21.09 0.6280 20.46 0.6238 19.43 0.6218
SRCNN 28 27.58 0.8990 20.97 0.9251 28.89 0.9161 29.26 0.9095
FSRCNN 29 29.83 0.9262 31.64 0.9420 30.67 0.9358 31.12 0.9302
VDSR 30 29.74 0.9205 32.70 0.9453 31.48 0.9387 31.72 0.9342
RCAN (CA) 32 31.52 0.9034 33.48 0.9267 32.97 0.9211 32.90 0.9132
41
RCAN (CBAM ) 34.52 0.9524 36.20 0.9611 35.68 0.9574 35.53 0.9511
RCAN (SENet ) 33.26 0.9363 35.31 0.9524 34.73 0.9484 34.75 0.9416
42
U-Net branch 34.09 0.9543 35.78 0.9622 35.07 0.9579 34.98 0.9506
RCEN branch 34.85 0.9597 36.48 0.9654 35.80 0.9620 35.44 0.9570
MPSR-Net 35.05 0.9561 37.17 0.9681 36.35 0.9645 35.94 0.9595
Abbreviations: CA: Channel attention; CBAM: Convolutional block attention module; EDSR: Enhanced deep super-resolution network; FSRCNN: Fast
super-resolution convolutional neural network; MPSR-Net: Melt pool super-resolution network; PSNR: Peak signal-to-noise ratio; RCAN: Residual
channel attention network; RCEN: Residual channel with the efficient channel attention network; SENet: Squeeze-and-excitation networks;
SR: Super-resolution; SRCNN: Super-resolution convolutional neural network; SSIM: Structural similarity index measure; VDSR: Very deep
super-resolution network.
Volume 3 Issue 4 (2024) 9 doi: 10.36922/msam.5585