<![CDATA[Road cracks significantly degrade infrastructure quality and pose a threat to traffic safety. To minimize manual inspection inefficiencies, this study investigates a segmentation model integrating MobileNetV3-Small as a backbone for the U-Net architecture to reduce processing time. The performance of the proposed MobileNetV3-Small-U-Net is benchmarked against a standard U-Net using three public datasets: DeepCrack (537 images), CFD (118 images), and Crack500 (3368 images) sourced from GitHub and Kaggle. This research explores the influence of optimization algorithms on evaluation results across these diverse datasets. Specifically, the study evaluates Adam, RMSprop, and SGD optimizers at an image resolution of 224 x 224 pixels, with a 0.001 learning rate and 0.9 momentum. On-the-fly augmentation techniques, including horizontal flips and brightness adjustments (0.8 to 1.2), were implemented during training. Experimental results demonstrate that MobileNetV3-Small-U-Net enhances computational efficiency by achieving a 9 ms inference time, which is 2 ms faster than the standard U-Net. These findings confirm that a MobileNetV3-Small backbone accelerates inference, despite a slight trade-off in evaluation metrics. Additionally, results reveal that the SGD optimizer is unsuitable for these segmentation tasks due to high error rates and the lack of an adaptive learning rate.]]>
Copyrights © 2026
