Design And Implementation Of An Embedded Multi-biometric Platform

As an important biometric identification technology, Face recognition and irisrecognition has been a hot spot in the field o f computer vision and pattern recognitionresearch. Because of its non-contact, safety and convenience, face recognition s widelyused in human-computer interaction, transaction authentication and security fields. Irisrecognition has been called as the "most accurate","the hardest fake" biometric technologybecause its high stability, uniqueness and non-invasive. Compared with iris recognition,face recognition is more direct and friendly, high user acceptance, but lower recognitionaccuracy; compared to face recognition, iris recognition has higher accuracy, but theacquisition is difficult. Inorder to achieve better recognition results, we can combine thesetwo features to make up their own weaknesses. Meanwhile, with the widespread use of therapid development of embedded systems and biometric identification technology, thedevelopment of embedded biometric technology, has become a hot topic in this field.First, we design the dual-core architecture for embedded multi-biometric identificationsystem based on DSP and ARM. We ake advantage of the DSP’s strong computing power,the face detection and face recognition and eye location algorithm is ported to DSP borad.The ARM runs the embedded Linux operating system, we write device driver of HPI hostinterface to transfer the facial features and the images from DSP’s stroge device to ARM.The use of DSP and ARM dual-core architecture enhances the interactivity of the entireembedded systems.Then, we use the Adaboost face detection algorithm based on Haar features to detectthe human face in the scene. Combined the results of face detection with prior knowledgeand grayscale integral projection algorithm, we locate the position of the human eye. Theimproved LBP feature extraction algorithm is used for face image feature extraction, andKNN algorithm is used for classification. As the algorithms of iris feature extraction is hardto achieve, this paper only collect iris images. We use the ORL face database to verify thecorrelation algorithm and port the program to embedded systems. To improve the detectionrate of the algorithm in the embedded platform, the optimization algorithms are used,including floating-point turn point, parameter optimization and memory optimization.Finally, based on our own face database captured at indoor scenes, the factors thataffect system performance is analyzed to determine the parameters of each part of the system. Performance of the system is tested, a large number of experiments show that thesystem is low power consumption, small, stable, fast and high accuracy which meet therequirements of real-time face recognition.