FACE IMAGE DE-OCCLUSION WITH VARIABLE-THRESHOLD ROBUST PCA

Abstract

The human face is one of the most important biometric traits in human social activities such as security, surveillance, identification, criminal and forensic investigation. This thesis aims to overcome the difficulties of existing methods for face occlusion removal by the proposed variable-threshold RPCA (VRPCA) method. Given prior knowledge of the occluded area, VRPCA separates the error matrix into three different parts: the training part, the testing unoccluded part and the testing occluded part. By applying different thresholds to these three parts, VRPCA ensures that the errors are mostly from the occluded part and that unintended corruptions is minimized in the reconstruction of the unoccluded parts.

Various test results show that VRPCA is consistently more accurate than existing methods across various test conditions for recovering occluded parts. Moreover, VRPCA is able to preserve the unoccluded parts of the target image with practically zero error.