Chinese character stroke extraction as a foundational task for various applications, including calligraphy analysis, aesthetic evaluation, and font identification. However, existing methods are often plagued by issues such as segmentation ambiguity, poor adaptability to structural rules, and high computational costs. To address these limitations, this paper proposes a novel method for Chinese character stroke extraction, termed UStroke, which incorporates prior knowledge into the process. Within an enhanced fully convolutional network architecture, a Feature-wise Linear Modulation (FiLM) layer is introduced to integrate stroke category prior knowledge, thereby strengthening the model’s responsiveness to specific stroke categories. Additionally, transposed convolution is utilized for up-sampling to enhance the representation of complex Chinese character structures. Experimental results demonstrate that on the LTH and FZLBJW datasets, UStroke achieves MIoU and FFWIoU scores of 92.25% and 93.35%, respectively, marking an 8% improvement over baseline models. Moreover, the inference time per image is only 120 ms. This method provides a high-accuracy and high-efficiency solution for Chinese character structure analysis.