PSI - Issue 74

Tomáš Vražina et al. / Procedia Structural Integrity 74 (2025) 106 –113 Tomáš Vražina / Structural Integrity Procedia 00 (202 5 ) 000 – 000

109

4

Melt pool tracks visible in Fig. 1a represent regions where the material was locally melted during the processing of individual layers. These tracks encompass several grains, within which near-hexagonal cellular structures are formed, as shown in Fig. 1b. The boundaries of melt pools often serve as sites for irregularities in material flow. A higher magnification STEM micrograph (Fig. 1c) reveals that these cells consist of dislocation networks. Dislocation rich cell structures are attributed to the strengthening of 316L steel. Proper characterization of the dislocation substructures and subsequent correlation with mechanical properties is therefore of great importance. Three different measuring methods were utilized to assess the characteristics of these cell parameters. Namely: a) the line-intercept method b) the adaptive thresholding, and c) the deep-learning U-net method. The diagram in Fig. 2 presents a hybrid segmentation workflow combining image preprocessing with a deep learning model based on a U-Net architecture.

Fig. 2. Simplified diagram of the hybrid workflow for segmenting dislocation cell structures, combining image preprocessing techniques with a U-Net architecture using a ResNet-based encoder. (ReLU – Rectified Linear Unit).

The input to the U-Net model consists of cropped images and microscopy image masks that are generated through a preprocessing pipeline that includes CLAHE (Contrast Limited Adaptive Histogram Equalization), adaptive thresholding, and subsequent morphological post-processing. The processed images are then passed through a modified U-Net network that uses a ResNet34 (He et al., 2015) encoder pretrained on ImageNet (database). Instead of max-pooling, the encoder employs stride convolutions to extract hierarchical features, which are later upsampled through transposed convolutions in the decoder. Skip connections bridge the encoder and decoder stages, preserving spatial detail for accurate segmentation. A final sigmoid activation produces a binary segmentation mask that identifies the dislocation cell regions. The dataset was divided into three subsets for model development and evaluation: 5–10 images were used for training, 2 for validation during model tuning, and 2 for final testing of segmentation performance. The U-Net model was trained for segmentation using the Adam optimizer with a learning rate set to 0.001. Training was conducted for 50 epochs, and binary cross-entropy loss was employed as the loss function. Data was processed in batches, and model performance was monitored through periodic validation loss assessment. 3. Results and discussion 3.1 Image thresholding and classification The etched microstructure in Fig. 3a was segmented using an adaptive thresholding method. The resulting cellular structures were then classified with the BIRCH (Zhang et al., 1996) (Balanced Iterative Reducing and Clustering using