Deep Corrosion: Semantic Segmentation Unveiled Through Advanced Learning

In this study, we address the underexplored application of deep learning in corrosion detection and assessment, utilizing cutting-edge models for semantic segmentation. Prominent aspects of this work include extensive use of diverse datasets, adapted for image segmentation, enhancing consequently the robustness of the selected models. In our investigation of loss functions, we contrast Dice loss with Focal loss. The achieved Pixel Accuracy is up to 97.38%, while the Intersection over Union (IoU) approximates 79.83%. The numerical outcomes reveal the superior performance of Focal loss, highlighting its substantial impact on corrosion image analysis, particularly in multi-class settings. This study spearheads the application of advanced neural networks in corrosion assessment, showcasing exceptional performance across diverse datasets. This achievement opens avenues for continued exploration in this field.