A Comprehensive Comparative Study on Semantic Segmentation for Automated Microstructural Measurement in Al–Si Alloys

This work presents an assessment of the performance of five deep learning segmentation models (DeepLabV3, FCN, LinkNet, SegNet, and U-Net) as tested on the microstructure of various Al–Si alloys with the goal of providing a more objective visual inspection of microscopic images. A variety of metrics, including Pixel Accuracy, Sensitivity, Recall, Dice Coefficient, IoU, and Specificity, are used for a comprehensive examination. The study addresses also object level mistakes and examines the performance of the models built for various Al–Si alloy microstructures, such as eutectic, hypoeutectic, and hypereutectic Al–Si alloys. Among the evaluated models, excluding SegNet, all delivered high and similar performance metrics. However, differences became more pronounced in post-segmentation contour numbers and average areas. The eutectic dataset had the lowest error rates at 4.2% for object count and 8.1% for average area. The hypoeutectic dataset showed even lower error rates with FCN, at 2.6% and 0.9%, respectively. For the multilabel hypereutectic dataset, while error rates for eutectic Si particles were similar, the DeepLabV3 model achieved the lowest errors in Flake Si measurements, with 0% in object count and 3.1% in average area.