The physiotherapy robot can provide personalized massage therapy by adapting to individual characteristics. The acupoint recognition serves as a crucial component within the physiotherapy robot system, wherein its deep learning-based acupoint recognition method demonstrates superior precision in identification. However, during dataset creation, directly labeling images poses challenges such as partial inability to annotate and visual errors. An indirect annotation method is proposed to address these challenges. The method entails manually marking points on the skin, followed by software-based annotation using the manual markings, and ultimately removing the manual markings from the images. The crucial component of this method is the de-marking algorithm, refined based on Criminisi. The improved algorithm improves the Criminisi algorithm’s priority calculation mechanism, establishes an equivalent elliptical model for regional directional matching. It addresses issues such as the failure of the original algorithm’s priority mechanism, low matching efficiency, and sudden pixel value changes after filling. The experimental comparison of the two algorithms shows the improved algorithm has better restoration quality and restoration efficiency. The result demonstrates that the indirect annotation method can enhance the quality and annotation efficiency of acupoint datasets, providing an effective method for annotating datasets for the acupoint recognition system of physiotherapy robots.