Research On Hyperspectral Face Recognition Based On Adaptive Band Selection

With the continuous development of machine learning and deep learning,face recognition as a biometric technology has entered all areas of people's lives.Hyperspectral face image is continuous on the continuous spectrum in a narrow band imaging formation of high-dimensional data,because the facial melanin in the skin under the different wavelength,hemoglobin and so on material has different reflection strength,make the hyperspectral face image compared to traditional biological characteristics of face image has more abundant,thus to solve the problem of illumination,anti-counterfeiting,posture,etc.However,the hyperspectral face image also brings many challenges,such as large amount of computation and redundant information features.Especially in the spectral dimension,the contribution of different wavelengths of wave band images to face recognition is different,and some even produce interference,which reduces the robustness of recognition.Therefore,band selection in spectral dimension is of great significance to improve the performance of spectral face recognition.In this paper,the research of adaptive band selection hyperspectral face recognition based on deep learning is mainly carried out.The main research work and results are as follows:(1)Based on the features of block and Adaboost,a hyperspectral face recognition method based on block Adaboost.MS band selection is proposed.This method divides the face into different regions,uses the convolutional neural network(CNN)to extract the face features of different regions and bands,and then uses the Adaboost.MS method to select the band,and then classifies the optimal band of the local region of the image by Adaboost.MS.Finally,the recognition results of different regions are classified using the maximum voting algorithm.To verify the effectiveness of the method.Experimental results of this method on Polyu-HSFD show that compared with other band selection methods,the accuracy of this method for hyperspectral face recognition is improved by 10%-13%.And the optimal band of different local regions of the face is different.(2)Based on the redistribution of weight of attention mechanism in the domain of image channels,a sparse spectral attention mechanism deep hyperspectral face recognition network(SSCANET)is proposed.The network is divided into two parts:band selection module and classification module.In the band selection module,the attention mechanism neural network module based on regularization(Lasso)is used to weight the different bands of the hyperspectral face image to achieve band selection.Then VGG-19 network module is used for classification.In order to verify the robustness of the proposed method,experiments were carried out on Poly U-HSFD,University of Western Australia Hyperspectral Face Database(UWA-HSFD)and Carnegie Mellon University Hyperspectral Face Database(CMU-HSFD),respectively.The results show that,The band selection module in this algorithm can realize adaptive band selection for hyperspectral face images through attention mechanism,and the computational cost of the band selection module is negligible compared with the computational cost of the VGG-19 classification module.