1 Introduction
2 Materials and methods
2.1 Procedure for measuring the melt pools
2.2 Generation of the training data
2.3 Semantic segmentation of the melt pools with deep learning
-
Pixel accuracy:$$\frac{\sum_{i}{k}_{ii}}{\sum_{i}{m}_{i}}$$
-
Mean accuracy:$$\frac{1}{K}\sum_{i}\frac{{k}_{ii}}{{m}_{i}}$$
-
Mean intersection over union (mean IoU):$$\frac{1}{K}\sum_{i}\frac{{k}_{ii}}{{m}_{i}+\sum_{j}{k}_{ji}-{k}_{ii}}$$
-
Intersection over union for the class melt pool border \({\mathrm{IoU}}_{\mathrm{MP}}\), which is indexed by b:$$\frac{{k}_{b}}{{m}_{b}+\sum_{j}{k}_{jb}-{k}_{b}}$$
2.4 Instance segmentation of the melt pools with the watershed algorithm
2.5 Identifying the layers
3 Results and discussion
3.1 Evaluation of the semantic segmentation performance
Pixel accuracy (%) | Mean accuracy (%) | Mean IoU (%) | \({\mathrm{IoU}}_{\mathrm{MP}} \left(\%\right)\) | |
---|---|---|---|---|
Without augmentation | 91.10 | 74.54 | 62.90 | 38.83 |
With augmentation | 91.11 | 75.81 | 63.58 | 39.75 |
Ground truth class labels | ||||
---|---|---|---|---|
Mounting material | Material | Defect | Melt pool border | |
Predicted class labels | ||||
Mounting material | 9898641 99.12% | 49965 0.36% | 1695 1.45% | 275 0.01% |
Material | 80787 0.81% | 13101670 93.24% | 60056 51.25% | 1292703 47.61% |
Defect | 708 0.01% | 46914 0.33% | 52261 44.60% | 2067 0.08% |
Melt pool border | 536 0.01% | 853694 6.08% | 3174 2.71% | 1420124 53.30% |