Segmentation Studies on Al‐Si‐Mg Metallographic Images Using Various Different Deep Learning Algorithms and Loss Functions

ABSTRACT Segmenting metallographic pictures is being done in material science and related domains in order to detect the features within them. Therefore, it becomes crucial to find grains and secondary phase particles. It is necessary to label every pixel in order to obtain satisfactory segmentation outcomes. However, labeling takes a lot of time and is physically taxing. Therefore, in order to obtain higher performance, we have suggested a semi‐supervised deep learning technique in the current study that uses fewer labeled images. Other deep learning algorithms, such as Segnet, Resnet, and FCN, were compared with the Unet approach that was suggested. Additional comparisons have been made using the Dice score (0.85), IOU score (0.74), F1 (0.85), and recall (0.96) measures. Different loss functions were also compared, including binary, SS loss, and Tversky. Furthermore, the dataset was expanded, and these datasets were also subjected to result analysis. The trials show that, both numerically and qualitatively, the suggested approach can produce superior outcomes with fewer labeled photos.