Fusing Iris And Face Biometrics For Personal Identification On Mobile Devices

Mobile devices are ubiquitous in our daily life.Large amount of personal data are generally stored on mobile devices and mobile payment is convenient and popular nowa-days.Traditional personal identification approaches are based on passwords or personal identification numbers(PINs),which are vulnerable to spoof attacks and cumbersome to remember.The information security of mobile devices is facing severe challenges.It is significant to develop more reliable personal authentication systems for mobile devices.Biometrics attract more and more attention.Among them,iris has advantages of high sta-bility,uniqueness,and anti-counterfeiting,making it suitable for usage on mobile devices to strengthen the security of personal identification system.Compared with traditional iris recognition methods,iris recognition on mobile de-vices faces some new challenges.Owing to limitations of camera sensors on mobile de-vices,the acquired iris images usually have inferior quality,high noise,and low resolution.In addition,because of the limited computation and storage resources on mobile devices,fast iris recognition algorithms with compact models are required.To solve the above-mentioned problems,our work is based on the iris recognition framework and develops solutions from the following aspects:iris image acquisition,image preprocessing,and iris feature analysis.What's more,when obtaining iris images,the face,especially the peri-ocular region can be simultaneously acquired.Therefore,we also propose multi-modal fusion approaches that fuse iris with face and periocular region to boost the performance of personal identification on mobile devices.The main contributions are as follows:1.A new NIR mobile iris database is constructed.It totally contains 11,000 images captured from 630 Asians.It is the largest NIR mobile iris database as far as we know.We have made it publicly available,which is beneficial for future related studies.2.Image super-resolution methods are proposed to enhance iris image quality.We adopt two pixel level super-resolution approaches:Super-Resolution Convolutional Neural Networks(SRCNN)and SuperResolution Forests(SRF).We analyze and compare the performance of these two approaches.The effectiveness of image super-resolution methods for improving the accuracy rate of iris recognition on mobile devices is verified.3.Iris feature extraction approaches based on the deep convolutional neural networks(CNNs)are proposed.CNNs have merits of automatically learning a high-level and abstract representation of the input image.We first adopt the Pairwise CNNs model to learn pairwise features of a pair of iris images and explore the comple-mentary relationship with ordinal measures features.Afterwards,we utilize the maxout CNNs model to directly extract the feature of one input iris image.It is more efficient and robust than traditional features.Higher performance is achieved for personal identification on mobile devices.4.Multi-modal fusion methods are developed,which can break through limits of a single modality and combine each modality's merits.We first fuse the iris and face modalities at score level based on the weighted sum rule.Then,we fuse the iris and periocular modalities at feature level based on the CNNs and propose the adaptive weighted strategy.Multi-modal fusion approaches achieve obviously better results than unimodal biometrics.Moreover,the proposed methods require fewer storage spaces and computational resources,which are suitable for mobile devices usage.In summary,the thesis systematically and deeply studies iris recognition and multi-modal fusion on mobile devices.A number of attempts are made and significantly im-proved performance is achieved for personal identification on mobile devices.