Crack detection is a vital aspect of road surface monitoring, ensuring the safety and durability of infrastructure. This paper presents an enhanced methodology utilizing the U-Net deep learning model integrated with adaptive thresholding for efficient crack segmentation. The process begins with robust preprocessing, including dataset preparation, image standardization, and visualization of image-mask pairs, to ensure high-quality inputs. The U-Net model, with its encoder-decoder architecture, is optimized using binary cross-entropy loss and the Adam optimizer to achieve superior performance in segmentation tasks. Key contributions of this study include the incorporation of adaptive probability thresholds (p > 0.7, p > 0.8, p > 0.9) to balance precision and recall, and the implementation of custom callbacks for continuous evaluation during training. Experimental results highlight the model’s effectiveness across diverse road conditions, demonstrating its ability to segment cracks with high accuracy and robustness. This pipeline sets a benchmark for integrating adaptive segmentation techniques into deep learning-based road maintenance systems, contributing significantly to intelligent infrastructure monitoring and smart transportation technologies.