Deep Rotation Vector Feature Convolutional Neural Network And Its Application In Face Recognition

As for face recognition has the characteristics of non-contact,convenience and friendliness,many researchers have studied it deeply for a long time and achieved many important results.In recent years,with the improvement of hardware equipment performance,the recognition rate of face recognition technology based on convolutional neural network is getting higher and higher,which becomes increasingly widely used in practical applications with obvious effect and superior performance.However,under the unrestricted conditions,the performance of the algorithm is greatly affected by changes such as illumination,expression,occlusion and age,which requires further research.Therefore,this paper mainly studies the recognition of blurred face images caused by unqualified conditions in face recognition based on the traditional deep convolutional neural network.For the feature extraction of face,this paper proposes two method for feature extraction based on the principles of Trace transform and analog digital signal methods:1.Rotation delta modulation feature extraction method,which performs delta modulation coding on each vertical trace in the image to extract the feature values.All the feature values at one Angle of an image form a one-dimensional vector,and the feature values at all angles form a two-dimensional matrix.2.Rotation mean pulsation feature extraction method.This method extracts the feature values of all vertical traces on an image through the mean pulsation coding idea proposed in this paper,likewise,the eigenvalues at one Angle form a one-dimensional row vector,and the eigenvalues at all angles form a two-dimensional matrix.Aiming at the classification of face recognition,Because the convolution neural network extracts the bottom features layer by layer,then extracts the high dimensional abstract features,while the features extracted by the above two methods are still the bottom features,therefore,cascaded the two methods with the deep convolution neural network and adding one feature extraction layer before the first convolution layer.For the face recognition of blurred faces,considering the perceptual mode of human eyes is that the closer the object is perceived by the human eye,the clearer it is and the farther it is,the more blurred it will be based on the above-mentioned cascade convolution neural network,introducing the image multiscale method and adding a multiscale layer before the feature extraction layer.In order to verify the recognition performance of the proposed method and network model in blurred human face,the experiment of this paper are carried out on Yale,ORL and AR three face databases with clear face image for training and corresponding blurred face image for recognition,The results show that compared with the traditional deep convolutional neural network,under the same experimental conditions,the recognition rate has been greatly improved.