Accurate segmentation of human sperm images is vital for automated semen analysis and fertility diagnostics. While deep learning methods perform well on sperm head segmentation, identifying the slender, low-contrast tail remains challenging. To tackle this, we propose URS, a unified framework that integrates U-Net-based learning with morphology-aware region growing for complete sperm segmentation. The pipeline begins with bilateral filtering and histogram equalization to reduce noise and enhance contrast. A customized U-Net variant then localizes sperm heads effectively under limited data conditions. To bridge the head and tail stages, a geometric analysis module estimates the head–tail junction by identifying the boundary point farthest from the head centroid. Tail segmentation is achieved through an adaptive seed selection strategy, which iteratively expands the region based on skeleton extraction and directional continuity. This enables the model to handle tail discontinuities and overlaps without requiring extensive annotations. Experiments on a newly constructed dataset demonstrate that URS outperforms both classical and deep learning baselines in head and whole-sperm segmentation. Ablation and sensitivity studies further validate the robustness and contribution of each component.