Research On Periocular Recognition Algorithm For Mobile Terminal

As a result of the poor sharpness and the serious loss of texture of mobile iris images,the traditional iris recognition algorithm can't achieve the desired performance.In order to solve this problem,this paper proposes a mobile terminal periocular recognition algorithm based on the fusion of periocular features and iris features.The main work of this paper is as follows:Firstly,the traditional mobile iris recognition algorithm is improved,and a mobile terminal iris recognition algorithm based on feature fusion is proposed.The complementary 2D Gabor feature and OM feature are selected by adjusting a large number of parameters.The similarity likelihood ratio is obtained from Joint Bayesian classification algorithm by combining the two features.When the sample ratio of training set and testing set is 9:1,our method achieves 1.2%equal error rate on the testing set of MIR-Train,which improves the classification performance of the algorithm.At the same time,in the aspect of periocular recognition of mobile terminal,the lightweight neural network Mobile v2 is optimized,and the FMSE module is proposed to reduce the feature redundancy of the anti-residuals module and enhance its feature expression ability.Using focal loss,the difficult samples are excavated to further improve the network classification ability.On the MIR-Train dataset,the FMSE module improves the EER by 1.2%,focal loss improves the EER by 0.77%,and the two optimization methods show some complementarity.In order to prove the effectiveness of the method,the optimized model is used to compare experiments on the UBIPR dataset.In this paper,the periocular recognition algorithm reaches 2.47%of EER on UBIPR dataset,which goes beyond the existing methods.Limited to the requirements of the mobile terminal for the size of the algorithm model,the optimized model is quantified.At the same time,the existing model compression algorithm is improved,and a model compression algorithm based on relative position storage of row index is proposed.After compressing,model is reduced from the original 11.15M to 1.90M,the EER is only increased by 0.77%and 0.22%on the MIRTrain and UBIPR datasets respectively,and the EER of the periocular recognition algorithm on the UBIPR dataset reaches 2.69%.Finally,a mobile terminal periocular recognition algorithm combining iris feature and periocular feature is proposed.The 2D Gabor feature of iris expansion images and the FMSE Mobile v2 depth feature of the periocular region are extracted respectively.The 2D Gabor feature simulates the characteristics of shallow networks,fusing with the deep network features of the periocular region,which realizes the complementary features.And then we use the Joint Bayesian to classify the fusion feature.When the sample proportion of training set and testing set is 5:2,the overall method reaches 1.83%of the equal error in the test set of MIR-Train.