Orientation Adaptive Response Order Pooling Based Gabor Feature Extraction And Fusion

Gabor filter has been widely used in palmprint recognition and face recognition.However,most existing methods fail to utilize texture information of all orientations when performing Gabor transformation.On the other hand,local feature pattern has been widely used in local texture description.We argue that it can extract more texture information from Gabor response to design an orientated local feature pattern,enhancing the robustness of Gabor filter based methods.This thesis firstly proposes an Orientation Adaptive Response Order Pooling(OAROP)method.In OAROP,the local pattern of each pixel of the 2D Gabor response is adaptively coded according to the orientation,and then the codes of all scales and orientations are pooled for feature description.The distance among the proposed features is measured by L1 or L2 norm.We perform extensive experiment of palmprint verification and recognition on four public databases,including contactless database,3D database,multi-spectral database,and hyper-spectral database.Experiment results show that the proposed method outperforms many popular methods on some aspects.For the verification task,the proposed method can boost up the equal error rate of the best result.For the recognition task,the error rate can be reduced to 0%.Those results show the robustness of the proposed method,and that the proposed method provides a competitive choice for palmprint recognition.Single biometric feature can only represent partial identical information.By contrast,fusion of multi-feature can utilize more identical information,enhancing the accuracy of the recognition.We perform feature fusion on a multi-spectral palmprint database,where the image information of one wavelength is viewed as one feature.We extract the OAROP features of each wavelength,and perform the score fusion of all the features of various wavelengths,having obtained better results than some popular methods.Compared with 2D Gabor filter,3D Gabor filter can explore more space information with the information of on the extra dimension.Most existing methods processing hyperspectral data are 2D based,having not fully explored the vertical information.This thesis extents the local feature pattern of the proposed OAROP to 3D form,enhancing the representation of 3D Gabor feature information.This thesis also proposes the feature fusion method of joint sparse representation classification,to tackle the problem of dimension cruse of 3D feature data.Several feature information obtained by 3D Gabor filter are jointly sparsely represented.Experimental results on the database of hyperspectral palmprint show that the 3D Gabor information gets effectively improved,achieving higher accuracy than other state-of-the-art methods.