1) Objectives: The current facial recognition tools are inefficient in predicting landmarks for facial palsy patients. Noticeable asymmetry in the face results in inaccurate results as the prediction models are trained on symmetrical faces. In this study, a method is proposed which takes advantage of the existing powerful machine learning tools which are trained on datasets of healthy subjects with symmetric facial movements to create a system that can analyze and localize facial landmarks on both healthy as well as facial palsy subjects.

2) Methods: The task is accomplished by a simple image processing algorithm where two symmetric faces are generated from a non-symmetric face image representing the left and right sides of the original image. This method was tested against two other methods. One, which uses the cascade of regression trees (CRT) algorithm and the other which is a retrained version of the CRT algorithm on a dataset of facial palsy cases called Massachusetts Eye and Ear database and model (MEE).

3) Results: The methods were compared on 3 different types of test datasets containing a total 125 images. The proposed method outperforms other two methods in cases of asymmetrical faces from healthy people and palsy patients with approximately 7% lesser error compared to the CRT method and 39% lesser error than the MEE method.

4) Conclusion: The proposed method had a considerably better performance compared to the other two methods, which opens new perspectives to address the problem of face landmarks localization problem on facial palsy cases.