Effectiveness of Automatic Segmentation with Deep Learning in Separating Overlapping Sheet Metal Parts

In manufacturing and quality control processes, the automatic segmentation of sheet metal components plays a critical role. However, overlapping and closely positioned metal parts pose significant challenges for traditional image processing techniques. This study evaluates the effectiveness of deep learning-based segmentation methods in distinguishing and accurately segmenting stacked or adjacent metal components. To compare their capabilities in identifying such configurations, both semantic and instance segmentation models were investigated. The models were trained using U-Net and YOLOv8 architectures, and their performances were evaluated using standard metrics including intersection over union (IoU), accuracy, sensitivity, specificity, precision, and F1 score. The IoU value for semantic segmentation was 0.9669, whereas for instance segmentation it was $\mathbf{0. 9 6 4 4}$. Experimental results demonstrate that the YOLOv8 model is capable of effectively distinguishing overlapping and closely situated metal parts, achieving higher accuracy in such scenarios compared to semantic segmentation approaches. These findings highlight the potential of deep learning-based segmentation techniques for sheet applications such as automatic identification and feature extraction of metal components, contributing to more efficient quality control, operational efficiency, part identification, classification, and production workflows.