Online Multispectral Palmprint Recognition
David Zhang,Zhenhua Guo
2010-01-01
Abstract:Palmprint, as a new member of biometrics family, has attracted much of research attention in the past decade. Many different algorithms and systems have been proposed and built. Although, great success has been achieved in palmprint research, palmprint recognition could be further improved in two aspects: higher accuracy and robustness against spoof attack. Multispectral imaging, a method to collect a series of images by different spectra, is a good technique to address the issues mentioned above. In this thesis, different aspects of multispectral palmprint recognition were investigated and discussed. First, the issues of multispectral palmprint recognition were addressed. A well developed and accurate multispectral palmprint prototype with high speed was proposed. There are four different kinds of illumination in the device developed, blue illumination, green illumination, red illumination, and near infrared illumination. The former three are the primary colors well known in visible spectrum, and they could compose other lights. A large multispectral palmprint database was built by this device. Multispectral palmprint recognition can be regarded as a kind of special multimodal biometrics and there are three popular schemes regarding multimodal fusion: image level, feature level and matching score level. In this thesis, the three levels of fusion were studied and compared. The shortcomings and advantages of each method make them applicable for different applications. Then, a critical issue for palmprint recognition was studied. Illumination which is used to enhance the palmprint feature is a key component in palmprint recognition system design. Although there are some rules or guidances on the selection of cameras, light types, lens etc. There is no work systematically evaluating whether the white light source, which is the dominant light color in palmprint recognition, and which is the optimal light. Based on the multispectral palmprint image acquisition and additive color theory, seven kinds of palmprint images are acquired by red, green, blue, cyan, yellow, magenta and white lighting. The question which light is optimal for palmprint was studied empirically through three kinds of palmprint recognition algorithms.After that, although a multispectral palmprint acquisition device was developed, the underlying design principles were not well studied. In general, more feature bands could provide more features and thus get higher accuracy. However, more feature bands may contain redundant information and require too much cost on computation. Thus, the optimal number of feature bands in terms of accuracy and computation cost is a key issue in multispectral imaging. In this thesis, a feature band clustering algorithm was proposed to determine the optimal feature band number. After determining the number, an exhaustive searching could find the optimal combination.Finally, CompCode, one of state-of-the-art algorithms for palmprint recognition was analyzed and used in this thesis. It is the first algorithm for extracting orientation information for palmprint image with good accuracy and less computation cost on matching. It is mainly composed of three parts: filter design, feature extraction and matching. Although some work has been done on proposing novel filters, little work has been done on feature extraction and matching. A novel feature extraction scheme, Binary Orientation Co-occurrence Vector, was proposed in this thesis. It showed robustness to rotation effect and got better results than CompCode did on public databases. There are two widely used distances for fast orientation feature comparison. They are SUM_XOR and OR_XOR. No one empirically analyzed which one is more appropriate for palmprint recognition, and their relationship is left open for analyzing. In this thesis, a unified distance measurement was proposed. There is one parameter to control the proposed distance, and SUM_XOR and OR_XOR are special cases of theproposed distance with suitable parameter value. It also empirically showed when a suitable parameter was selected, better accuracy could be achieved comparing with the two distances.