Laser-directed energy deposition (LDED) has emerged as a promising metal additive manufacturing technique owing to its small heat-affected zone, low material wastage, and minimal environmental constraints. It has been widely applied in aerospace component fabrication, brake disk coating, and high-value part repair. However, the quality and stability of LDED are highly dependent on process parameters such as laser power, powder feed rate, and scanning speed. Conventional parameter optimization largely relies on post-build inspections, in which the manual measurement of bead morphology and cross-sectional geometry is both time-consuming and error-prone. To address these drawbacks, in this study, we propose a deep-learning-based geometric data measurement framework for weld bead characterization. A U-Net-based semantic segmentation model was developed to analyze both appearance morphology and cross-sectional geometry in single-track deposition experiments. The proposed models achieved feature recognition accuracies exceeding 90% on test datasets, demonstrating robust inference performance. Furthermore, the integration of automated image-based inspection software reduced measurement time by approximately 91% compared with manual evaluation. The proposed framework demonstrates the potential of deep-learning-assisted image analysis to improve process analysis efficiency, reduce human-induced errors, and enhance parameter optimization in LDED, thereby contributing to more reliable and intelligent additive manufacturing systems.