Research On Face Recognition Algorithm Based On Dilation Convolution

Face recognition technology has undergone significant changes in recent years.Traditional methods combine manually designed texture features with machine learning technology,making it difficult to extract features from different situations in an unconstrained environment.In past research,researchers have focused on researching methods for specific changes.Deep learning has taken over traditional face recognition methods.The advantage of deep learning methods that learn from a large amount of data is that the dataset contains various facial features.In this case,the trained model can learn the best features to represent these data.This article studies facial recognition methods to reduce the number of parameters while improving facial recognition rate.The main content and research results of this article are as follows:To address the issue of overfitting in residual network training,a residual network algorithm that introduces hollow convolution is proposed.The hole convolution is introduced into the residual module,and the Receptive field is expanded by using the hole convolution with the expansion coefficient of 2 in the residual block with the step size of 2.The information is fused through the residual connection.At the same time,the batch normalization layer is used to accelerate the Rate of convergence,and the accuracy rate of 99.48% is achieved in the test of LFW dataset.In response to the current problem of complex facial recognition network model training that requires a large amount of facial data,an algorithm based on the DInception network model is proposed.Pass 1 × 1 Convolution increases the nonlinear characteristics of the network,extracts multi-scale facial feature information with the Inception module,expands the Receptive field by introducing cavity convolution in the Inception module,sparsely samples,uses the global average pooling layer instead of the traditional full connection layer,reduces network parameters,and finally uses the joint loss to reduce the within class distance of the same feature.After training with small data sets,the accuracy rate is 97.58%.In order to improve the recognition rate of Convolutional neural network,the number of network parameters is too large,resulting in a long training time.A depthwise separable convolution algorithm based on bottleneck layer structure is proposed,which reduces the expansion coefficient of the Bottleneck structure in the Mobilenet V2 network and adds a depthwise separable convolution.The inverse residual method is used to extract high-dimensional facial features and improve accuracy.The use of an improved Bottleneck structure deepened the network layers,reduced the number of parameters,and achieved an accuracy of 99.32%.