This report explains what was done and learned by a project whose major goal was to research and develop a forensic tool with the capability of performing facial recognition, using low-quality, low-resolution faces, such as those obtained from closed-circuit television surveillance footage of crimes in progress.
A major achievement of the project has been the development of a "unified face representation model" that can interpret face-degradation scenarios, such as low resolution, pose, occlusions, etc. Thus, it has reinterpreted the problem of face recognition and recovery under acquisition degradation as a missing-data recovery problem. The developed method can also be used as a pre-processing step to aid in recognition by several commercial face recognition engines, thereby expanding the scope of these tools. Using this representation of faces, the project team was able to develop an automated facial occlusion-recovery system. This system can reconstruct parts of the face that are not visible in an image due to an obstruction. Possible obstructions include scarves, masks, sunglasses, eyeglasses, hair, etc. In addition, the project developed a method for recovering both the representation vector and a set of confidences for off-angle face images. Using these techniques, along with data-completion techniques described in previous reports, frontal facial images can be reconstructed and passed through face-recognition engines. The project also developed a periocular reconstruction and recognition technique. This technique recovers full-face images based on just the periocular region of the subject. The recovered full face can then be used for face recognition, thereby overcoming the limits of matching subjects wearing masks or burkas. This report also describes how project results have been disseminated to communities of interest. 41 figures and 1 table