This paper addresses the limitations of existing face keypoint detection algorithms when applied to cinematic and television contexts. In response to the challenges posed by occlusion and heavy makeup, which often lead to inaccurate face keypoint detection, we have developed a film and television face dataset comprising over 10,000 images. This paper introduces a hybrid model detection framework that combines PFLD_GhostOne and B-Spline curve for enhanced accuracy. In our methodology, we extract face contour coordinates from both the PFLD_GhostOne and dlib68-point face keypoint detection models, and establish a more precise skin contour by computing the distance between corresponding points via uniform cubic B-Spline curve interpolation fitting. Experimental results indicate that this approach provides robust detection accuracy for face keypoints in cinematic and television content, whilst maintaining efficient algorithmic speed.