Finger-Knuckle-Print Recognition Based On Feature Coding And Regression Analysis

Feature extraction is one of the key points in the field of pattern recognition. The aim of it is to extract the effective discriminative information from the samples for pattern classification purpose. Feature coding and regression analysis methods are widely used for feature extraction due to their favorable properties, such as low computational cost and effectiveness. In this dissertation, we develop some novel feature extraction techniques for finger-knuckle-print recognition. Furthermore, we compare these proposed approaches with the previous popular algorithms and validate the effectiveness of our approaches. The main work and innovation of this dissertation are included as follows:(1) A reconstruction and adaptive binary fusion (R-ABF) method is proposed. Duo to the fact that the collected images may have some pose variations, R-ABF tries to alleviate the effect of the pose variations from two aspects:one is to reconstruct the images which may have pose variations using the template images, aiming at reducing the enlarged distance caused by pose variations; the other is to develop an adaptive binary fusion scheme from the idea of multi-model biometrics, aiming at considering the effect of both the samples before and after reconstruction. Based on above two considerations, the proposed R-ABF algorithm can reduce the false rejection rates to some extent without increasing much the false acceptance rates, resulting more stable and accuracy recognition system.(2) A multiple orientation and texture coding integration (MoriCode&MtexCode) scheme is proposed. Due to the fact that only one dominant orientation is extracted in the previous methods, we propose to code each filtering response by using the multiple characters of the Gabor filters. Meanwhile, we try to extract the local texture information using local binary pattern (LBP) under the multiple orientation framework. Furthermore, the possibility of each bit having dominant orientation is depicted and incorporated into the distance matching procedure. Finally, multiple orientation and texture information are integrated via weighted average fusion scheme. Extensive experimental results also verify the effectiveness of the proposed MoriCode&MtexCode scheme.(3) A weighted competitive coding (W-CompCode) method is proposed. Based on the competitive coding (CompCode) scheme, we utilize the variation of Gabor filtering response for each pixel to design a weight and incorporate them into the conventional distance matching procedure. Meanwhile, we code the designed weight and also perform distance matching on them. The weight pattern distance is finally fused with the corrected matching distance via score level fusion scheme. Compared with CompCode, the proposed W-CompCode can improve the recognition performance a lot.(4) A Bayesian Appearance Regression (BAR) model is proposed. In visual classification tasks, the appearance of the training sample images also conveys important discriminative information in addition to the class label. The proposed BAR can simultaneously exploits the sample class label and the sample appearance to set the value of the regression response based on the Bayesian formula. BAR learns a linear mapping using ridge regression to extract the image features, and the classification can be simply done by the nearest neighbor classifier. Compared with the previous methods, the proposed BAR has advantages such as small number of mappings (equals to the number of classes), small feature storage space, fast feature extraction, insensitiveness to input feature dimensionality, and robustness to small sample size.