Stand-off Face and Iris Recognition in Unconstrained Environments.

摘要

With the increase of security concerns from industry and government, biometrics provides solutions for secure identifications and verifications. Most successful applications in this field have mainly focused on controlled conditions and cooperative behaviors of subjects, utilizing knowledge from databases that were collected in controlled and ideal conditions. It is of great interest and increasing need for measuring biometrics in more complicated scenarios, e.g., with less information, rapidly changing environments, limited resources, unpredictable behaviors, etc., with the minimal participation from the target person. However, performance degradations are seen in biometric systems due to imaging conditions (e.g. illumination variations, resolution issues, noise, blur, occlusion, etc.) and the complex structure of biometric modality (e.g. aging, pose). Hence, understanding biometric data and the impact of different factors in recognition performances is critical for biometric systems. When and how to utilize different sources of information from target subjects in order to boost are becoming more important.;This work investigates several factors that can affect face and iris recognition performances in order to improve multimodal fusion performances. This thesis introduces an unconstrained multi-modal and multi-spectral dataset that can be utilized for understanding the benefits of recognition performances in unconstrained environments. The contributions of this work are four folds: First, a dynamic multi-modal fusion framework based on quality-selection scheme is proposed to select the quality factors for each modality that have major contributions to the system performance. A solution of utilizing a vector of selected quality measures is also proposed and examined in different fusion strategies to improve fusion performances.;Second, a unique data collection protocol of unconstrained behaviors is designed and utilized to collect a new dataset focused on unconstrained multi-spectral and multi-modal biometric data at a distance, titled as Q-FIRE II Unconstrained dataset, frontal face images with different facial expressions in multi-distances spanning visible, near infra-red and long-wave infra-red spectrums is extracted and organized as a face sub-set of Q-FIRE II Unconstrained dataset. Additionally, a fast and pose-invariant long-wave infra-red face detector is also proposed and trained based on long-wave infra-red face images. To understand the benefits of utilizing multi-spectral face data in non-ideal conditions, cross-expression face recognition performances are examined on visible and near infra-red data.;Third, the impact of out-of-focus blur on face recognition performance is studied. Eleven sharpness levels based on the modulation transfer function (MTF) quality measures are utilized to examine the impact of out-of-focus blur on face recognition performance based on a range of controlled real face out-of-focus blur during acquisition from the Q-FIRE dataset. The MTF method for measuring sharpness is proposed to compare with other sharpness measurements with a reference of the co-located optical chart next to the face region from the dataset. Three different face recognition systems are examined utilizing different sharpness level.;Lastly, to study the impact of time lapse on iris recognition, the adjustment of recognition performance based on quality factors and the statistic method of regression are introduced and incorporated in the investigation of the impact of time lapse on iris recognition.