Optimizer Influence on U-Net Architecture for Crack Segmentation

Crack detection and segmentation on surfaces are critical for maintaining structural integrity in various engineering domains. This paper presents an innovative approach to automated crack segmentation using a U-Net-based architecture, which was trained and tested with diverse optimizers to evaluate their impact on model performance. Our study compares six optimizers: Adam, RMSProp, Adamax, AdamW, Adadelta, and Adagrad, across various metrics including accuracy, loss, and the Dice and Tversky coefficients. The experimental results, obtained from a substantial dataset of surface images, suggest that while the Adam optimizer achieved the best balance of accuracy (97.25%) and loss (4.26%), AdamW yielded superior results in the Dice (68.50%) and Tversky (68.60%) indices, indicative of its robustness in crack delineation. The study also highlights the limitations of Adadelta, which underperformed significantly across all metrics. Focal Tversky index was introduced to weigh in the false negatives more than false positives, offering nuanced insights into segmentation precision. These findings contribute to the ongoing discourse in image segmentation and pave the way for more advanced crack detection systems that are vital for preemptive maintenance and safety assessments in civil infrastructure.