Study Of Lightweight Face Recognition Method Based On Attention Mechanisms

Face recognition is one of the most regularly utilized authentication methods at present time.Face recognition systems based on large-scale convolutional neural networks have been frequently developed and used as deep learning technology advances.However,these systems typically occupy a substantial amount of computational resources and cannot be identified in real time on mobile and embedded devices.As a result,lightweight research in face recognition and improving the accuracy of lightweight models have emerged as popular study directions in the field of face recognition at this point.As the attention mechanism continues to be proposed,it has become one of the most often utilized approaches to increase the accuracy of the model through the application of the attention mechanism at this stage.Based on this,this paper investigates and analyzes the use of the attention mechanism in the lightweight model,and it provides improved approaches.This paper's key contribution is as follows:1.The feature density is increased by introducing extended convolution into the channel attention mechanism.In view of the fact that dimensionality reduction in channel attention mechanism will destroy the direct correspondence between channel and its weight,onedimensional convolution is used in this paper to replace the two fully connected layers,so as to improve the feature extraction ability of channel attention and reduce the amount of computation.At the same time,a lightweight face recognition algorithm named Fast Face Net is proposed based on depth separable convolution.Experimental results show that the method is effective.2.To address the issue that the SENet only focuses on information exchange between channels and does not extract more image information,this paper employs the convolution module and the Transformer module in parallel to achieve the fusion of local and global features and to improve the model expression ability.Simultaneously,an adaptive penalty angle loss function is presented to increase the accuracy of the lightweight face recognition model.It is avoided that when the sample angle is tiny,the applied penalty angle is too big,resulting in the problem of missing the optimal solution during the training process by adaptively modifying the size of the penalty angle.When the model angle is bigger,a larger penalty angle is applied to accelerate the model's training speed.3.Aiming at the above approach,this paper proposes an AR-based teacher assistant.Using the deep learning method to identify students,achieve non-sensory attendance,and assist teachers in understanding student information and implementing individualized teaching.On mobile devices,the teacher assistance app can obtain 25 FPS.