Performance of Face Recognition Algorithms on Compressed Images

摘要

This report provides a comprehensive assessment of the ability of face recognition algorithms to compare compressed standard face images. Six well performing algorithms from the Multiple Biometric Evaluation (MBE) 2010 Still Face Track are used to compare face images compressed in JPEG and JPEG2000 formats. A primary goal of this report is to identify the maximum storage constraints under which verification systems can effectively operate. Toward this end, guidelines and recommendations are provided for the efficient compression and storage of face images for biometric applications. Many biometric systems operate with restrictions on the size of stored samples. For example, smart-cards (e.g. ePassports and government PIV cards) have a maximum storage capacity for storing their biometric samples. In addition, systems employing limited-bandwidth networks are often bottlenecked by the size of the biometric data they transfer. Although proprietary formats for storing facial features (i.e. matchable templates) are often more compact, their use should be avoided for several reasons. First, their content is non-standard, non-interoperable, and not suitable for cross-agency exchange of data. Second, their widespread adoption would undermine fair competition in the marketplace. Lastly, when the original image is discarded, the human interpretable representation of the face is lost. Discarding the original image in exchange for the proprietary template would also prevent systems from benefiting from inevitable future improvements in the automated feature extraction and template generation process.