Application Research On Gabor Wavelet Transformation For Face Recognition

Motivated by the extensive potential applications in public security, financial security, human–computer interaction, etc, face recognition has become one of the most active research areas in computer vision. Over the last decades, numerous various face recognition approaches have been developed and successfully applied in the real world. However, automatic face recognition is still a difficult problem in computer vision. The main reason for this is that the face recognition task is usually confronted with image variations in pose, facial expression, illumination condition or aging. So a practical face recognition technique needs to be robust to these variations.The Gabor filters, whose kernels are similar to the response of the two-dimensional receptive field profiles of the mammalian simple cortical cell, exhibit the desirable characteristics of spatial locality, spatial frequency and orientation selectivity, so often act as a powerful tool to extract the robust features from face images. However, the overwhelming high-dimensional Gabor feature has high cost in space requirement and makes recognition process computationally expensive. To reduce the dimensionality of the Gabor feature, the popular strategy is to down-sample the filtered images with a factor or employ the Adaboost algorithm to select the most discriminative Gabor features after convoluting the face image with Gabor kernels. Then the down-sampled or seletected Gabor features are concatenated to form a feature vector to represent a face image. Nevertheless, the down-sampling technique discards a great number of discriminative Gabor features, and the resulted feature vector still resides in a space of high dimensilonality. For the latter, it is very time-consuming to select the most useful ones from so many Gabor features. Besides, one disadvantage shared by them is that face matching requires alignment of corresponding pairs of pixels ideally, which is difficult to satisfy and then leads to mis-alignment.To address the above-mentioned problems, this paper explores Gabor texture features of face images in different views. Moreover, the two-dimensional manifold subspace method is applied to the Gabor magnitude. Generally, the main innovative fruits of this thesis include:①A novel approach, called the null sapce linear discriminant analysis (NLDA) of Gabor statistical texture features, is proposed for face recognition. The statistics (mean and standard deviation) of the Gabor magnitude taken as texture features, have been widely and successfully applied in content-based image retrieval or segmentation. They can make full use of the discriminability implied in all the Gabor features when the dimensionality is largely reduced. Therefore, this paper applies them to face recognition for the first time. To obtain more robust local texture features, in this paper the Gabor-filtered image is partitioned into several equally sized, non-overlapping sub-images, where the statistics such as mean and standard deviation are calculated and used as local texture representation. Finally, these local texture features of all the sub-images are concatenated to form a low-dimensional feature vector. Before being used for face recognition, the feature vector is subjected to NLDA to obtain enhanced discriminative power. Experimental results have shown that our method can effectively reduce the dimensionality of Gabor features and improve the recognition accuracy when compared with NLDA of the down-sampled Gabor features.②From the view point of psychological research on human texture perception, we employ the marginal densities of Gabor magnitude and phase as texture features, respectively, and propose two new kinds of texture representation methods for face recognition: Gabor magnitude-based texture representation (GMTR) and Gabor phase-based texture representation (GPTR). Firstly, the Kullback-Leibler distance (KLD) is used to validate that the distributions of the Gabor magnitude and phase can be modeled accurately by the Gamma distribution (ΓD) and the generalized Gaussian distribution (GGD), respectively. Then, GMTR is characterized by using theΓD to model the Gabor magnitude distribution, while GPTR is characterized by using the GGD to model the Gabor phase distribution. The estimated model parameters serve as texture representation. The proposed GMTR and GPTR can greatly reduce the dimensionality of Gabor features with a condensed representation of the face image. In addition, the fusion of them at feature level is also obtained to validate the complementariness of them. Our experimental results show that GMTR and GPTR significantly outperform the down-sampled Gabor features. In particular, the feature level fusion of them performs better than them individually.③Two new kinds of texture representation methods are introduced for face recognition: Gabor real part-based texture representation (GRTR) and Gabor imaginary part-based texture representation (GITR), which use the marginal densities of Gabor real part and imaginary part as texture features, respectively. The distributions of the Gabor coefficients, i.e., the Gabor real part and imaginary part, can be modeled accurately by the GGD, so GRTR and GITR are obtained using the GGD to model the real and imaginary parts, respectively. The estimated model parameters serve as texture representation. Similarly, the proposed GMTR and GPTR can greatly reduce the dimensionality of Gabor features with a condensed representation of the face image. To validate the complementariness of the Gabor real part and the Gabor imaginary part, the fusion of GRTR and GITR at matching score level is also obtained. Experimental reuslts demonstrate that both the proposed GRTR and GITR are superior to the widely used down-sampled Gabor features. Specially, the fusion of them at matching score level achieves better performance.④A new face recognition algorithm, named sub-pattern based two-dimensional locality preserving projections (Sp-2DLPP), is presented in this paper. 2DLPP fails to deal with well the sensitivity to large variations in head pose, lighting condition and facial expression since it only utilizes the global information of face images. In order to enhance the robustness of 2DLPP, we combine 2DLPP and the sub-pattern technique for the first time and propose the Sp-2DLPP method. Sp-2DLPP partitiones each face image into several equally sized, non-overlapping sub-pattern images, on which 2DLPP is performed to extract local features. Finally, a set of base classifiers are constructed based on different sub-pattern's features respectively, and then combined using the sum rule for the final decision. Therefore, Sp-2DLPP utilizes not only the local information of a face image but also the spatial structural information of each local region. The experimental results reveal that the proposed Sp-2DLPP can significantly improve the robustness of 2DLPP.⑤Sp-2DLPP based on the Gabor magnitude is developed for face recognition. As mentioned before, the conventional methods concatenate the down-sampled Gabor features to represent a face image, which results in a high-dimensional feature vector space. Sp-2DLPP directly operates on each sub-pattern image denoted as a sub-matrix partitioned from the original whole image, which does not need to transform the sub-image matrix into the same dimensional vector, so it can avoid the problem induced by high dimensiolity of Gabor features. Hence, this paper applies Sp-2DLPP to the Gabor magnitude. The experimental results demonstrate that Sp-2DLPP based on the Gabor magnitude performs better Sp-2DLPP based on the gray level features of face images obviously, which indicates the Gabor features carry more disciminative information than the gray level features do.