Research On Biological Characteristics Recognition Based On Deep Learning

Biological identification has become an integral part of today's society,and it has been applied to life,business and national security.This paper mainly studies the application of convolutional neural networks(CNN)in deep learning in the field of face recognition and voiceprint recognition.This paper has carried out a series of studies on the existing problems of deep learning in recognition and improved the existing algorithms.In this paper,the proposed algorithm reduces the training parameters and training time of the network without affecting the recognition accuracy,enhances the generalization of the network and reduces the possibility of over-fitting.The mail works of this paper are described as follows.(1)An improved pooling method is proposed by comparing several pooling methods of the pooling layer in convolutional neural network.This method uses the square probability distribution of the activation value to carry on the stochastic pool,which replaces the original maximum pooling method.By using this method,we can not only preserve the max pooling for image texture feature extraction,but also apply the advantage of stochastic pooling,which preserves the possibility of extracting hidden features in image and effectively enhances the generalization of the network.(2)Because the recognition image may be affected by various factors,the recognition effect is not good.Therefor,in view of the situation that the image is affected by illumination,small and poor quality,based on the image pretreatment of the image normalization and image demean value,this paper adopts the histogram equalization to reduce the illumination effect,and random clipping enlarges the number of images so as to reduce the possibility of network over-fitting,and uses the Gabor wavelet transform to enhance the images.Then,the Faster R-CNN network is used to perform the face detection on the LFW database.The existence of three fully connected layers of the traditional VGG-16 network will cause a large number of parameters in the network training.This paper reduces the number of fully connected layers and improves the random square number by improving the traditional VGG-16 network.The pooling method replaces the original maximum pooling method by referring to the GoogLeNet network method.The last pooling layer is changed to the global average pooling.The experiments are performed on the LFW database and the self-built database.It is found that the network training parameters are effectively reduced,the time of network training is greatly reduced and a good recognition rate is obtained.(3)In the voiceprint recognition,the speech signal is quantized and peremphasized,then it is processed by framing and adding window.The voiceprint information of the self-built digital voiceprint library is converted into spectrogram.The voiceprint recognition experiment is carried out on the LeNet-5 network using a grayscale digital spectrogram and on the VGG-16 network using a three-channel color digital spectrogram.Then an automatic speech recognition system was built to perform voiceprint recognition experiments on real-time speech information.