Face Recognition Technology Based On Channel Attention

Since the emergence of face recognition technology,through continuous research and development,the technology has gradually matured.The current mainstream face recognition technology is based on deep convolutional networks,and its models usually use smaller convolution kernels as image feature extraction tools.Small convolution kernels have fewer computing parameters and can save computing resources;under the same computing resources,models with small convolution kernels can achieve better model performance.Today,with the continuous improvement of the deep network structure and the increasing network depth,the performance improvement of the face recognition model encounters a bottleneck.Research has found that the performance of the model is closely related to whether the model can better extract data features.Through a deeper network model,the computational horizon can be expanded to extract multi-scale spatial features,but this method only improves the theoretical calculation horizon.The effective field of view is not large,and the extraction of channel features is rarely concerned.In order to effectively extract the spatial features and channel features of face images,this paper will use a large convolution kernel to increase the effective computational field of view,so as to extract spatial features more effectively.For the problem that the convolution kernel is too large and leads to too many parameters,structural re-parameterization is used to optimize the model inference;for the extraction of channel features,starting from the attention of the channel,a channel attention module is proposed based on the connection between channels,and Apply it to the face recognition model.Finally,the effectiveness of the proposed module is verified by ablation experiments.The main work of this paper is as follows:1.A face recognition method based on channel attention and enlarged field of view is proposed.To expand the model’s aggregation of spatial information,a larger convolution kernel is used for feature extraction on the input image.This paper selects Rep LKNet with large convolution kernel as the backbone structure of the model,and introduces a dual-branch enhanced channel attention module named Dual CAM.The input part of the model is adjusted to fit the resolution of the training images,and the global separable convolution(GDCon V)is used in the output part to reduce the parameters of the fully connected way.Batch normalization is widely used in the model,which effectively improves the feature extraction and parameter stability of the model.Structural reparameterization of the model in the model export phase can improve the inference speed of the model.The experimental results show that the proposed face recognition method has a high recognition accuracy.It should be emphasized that in the deep neural network structure,the convolution structure as the core of the model mostly adopts smaller convolution kernels.In this paper,the channel attention strategy with large convolution kernel model as the basic backbone network and structural reparameterization combined with dual branches is extremely rare in the field of face recognition.2.Aiming at the limitation of the hardware conditions of portable devices,the common face recognition model cannot be deployed and run well.A light-weight face recognition method based on channel attention is proposed.The face recognition method is improved on the basis of Mobile Face Net.From the perspective of paying attention to the channel connection,combined with the lightweight channel attention module Light CAG,it further promotes the feature interaction ability across channels and improves the feature extraction ability of the model.The experimental test shows that this method has a high recognition accuracy.