Hyperspectral Face Recognition With Patch-Based Low Rank Tensor Decomposition And PFFT Algorithm

As a biometric information,human face has been widely applied to the personal authentication,video surveillance,human-computer interaction,etc.Compared with other biometric information such as iris and fingerprint,human face has non-contact feature and can be recognized from a distance,but the recognition process is still challenged by unrestricted conditions(such as side faces,rich expressions,light and occlusion).With the development of hyperspectral acquisition system,it brings new opportunities to face recognition under nonrestrictive conditions.Compared to other forms of face images,hyperspectral images not only obtain spatial information but also additional spectral information,which realize the real combination of spectral information and image for the first time.With the development of face recognition technology,skin color segmentation,face feature extraction and other technologies,as important steps of face recognition technology,have attracted researchers' attention.In this paper,the spectral information of hyperspectral faces with small samples is fully explored and fused with spatial information to improve the recognition accuracy.The main work includes two aspects: skin color segmentation and feature extraction.In this paper,a new integrated method for skin feature segmentation of hyperspectral face images based on K-means clustering and minimum spanning forest classification algorithm is proposed,which makes use of both spectral and spatial discriminant features.According to the closed skin area,local features are selected for further facial image analysis.Firstly,the Kmeans is completed on a single band,and clustering results of different bands are compared to generate neighboring pixel blocks with tag information.The unclassified pixel blocks are then classified according to the similarity of the spectral information according to the minimum spanning forest algorithm.Compared with the traditional pixel-by-pixel segmentation methods,we use neighborhood uniformity to relabel the basic clustering results to reduce the computational complexity and obtain more accurate skin boundaries.Because the high homogeneity of adjacent regions of the image,the pixel value mutation occurs only at the edge of the facial organ,which has a relatively little impact on segmentation.In addition,this paper designs two feature extraction algorithms based on spectral information and spatial information of human face,namely the local feature extraction method based on block-based low-rank tensor decomposition and the global feature extraction methodbased on fast Polar Fourier transform.Some traditional image processing techniques regard images as vectors or matrices,which cannot make full use of spectral information,resulting in the loss of image information.Tensor is a powerful tool for processing and analyzing highdimensional data.It is an extension of vector and matrix data.The sparsity and orthogonality constraints are applied to tensor decomposition,that is,the four-dimensional tensor information composed of the same local regions of the face image is decomposed into a sparse coefficient tensor and three dictionary matrices to extract the detailed information of the face,which have better robustness for the non-restrictive environment.The global feature extracted in this paper is the contour of human face.In terms of face image,the gray value of detail information in local areas changes slowly,while the gray value of global contour information changes dramatically.The frequency domain of the Fourier transform represents the intensity of the grayscale transformation in the image,so this paper chooses the low-frequency information of the Fourier transform to represent the global contour feature of the face.The fast Polar Fourier transform extracts the overall information of the fused hyperspectral face image(that is,the facial contour information).Compared with traditional grayscale or color images,the fused two-dimensional hyperspectral images can not only increase the spectral information,but also can quickly organize and understand the rotation and scaling of images,reducing the risk of external factors.Finally,local feature classifier and global feature classifier are fused to obtain more robust and accurate classification results.