Key points are not available for this paper at this time.
Detection and evaluation of anomalies within a given human facial image is a challenging task. In particular, distinguishing between normal and abnormal attributes of the richly detailed human face is especially difficult. This work presents a face normalization approach, based on a novel automatic mask generation algorithm for image in-painting, called Auto-MAT. Auto-MAT helps solve the problem of facial anomaly detection by utilizing the concept of reconstruction-by-in-painting. Auto-MAT relies on sampling masks from a binomial distribution and selecting the optimal mask that identifies the anomalous facial component using a transformer-based in-painting method. Auto-MAT iteratively adjusts the randomly generated mask to detect the entire anomalous region and replace it with normal facial features. The effectiveness of the Auto-MAT is demonstrated through the successful processing of images depicting congenital cleft lip anomalies, and the generation of credible, normal whole-face analogues for the considered original images. Auto-MAT requires significantly less execution time than previously proposed facial anomaly normalization methods and has been incorporated into a facial appraisal platform, that corrects abnormal faces and provides an anomaly severity index that correlates closely with human ratings. Such a platform can be used as an effective clinical tool for the objective assessment of human facial anomalies and reconstructive surgical outcomes.
Hayajneh et al. (Mon,) studied this question.