Improving face verification in photo albums by combining facial recognition and metadata with cross-matching
MetadataShow full item record
Facial recognition is an important tool used by many disciplines, but its wider use in face detection and identification tasks has been somewhat limited. This is due to the many uncontrolled factors affecting faces in images, such as lighting, orientation, hair obscuration, blur, and the effects of aging. Despite tremendous efforts to overcome these uncontrolled factors, the reliability of a computer-based face recognizer is still questionable. In our research, we address the possibility of improving face verification using weighted cross-matching, which relies on a face verification metric and metadata. The idea is to implement a framework compatible with multiple platforms and capable of operating with limited resources while achieving satisfactory performance. We do not use statistical models, and we do not create patterns that require supervised learning. Our methodology is intended for use in personal digital image libraries because these libraries represent naturally context-correlated datasets. We use the native connection between files to determine the trustworthiness of an image relative to another. We then use this metric to attribute weights to pre-identified faces that are used as cues to help verify ambiguous elements. The final algorithm does not require the user's collaboration and performs automated digital image library management.
Approved for public release; distribution is unlimited