Face Recognition Of Small Samples Based On Transfer Learning In Specific Scenarios

In recent years,the rapid growth of data volume and the rapid development of computer hardware and software have provided important basic conditions for deep learning.Because deep learning requires a lot of data for training,in some specific face recognition tasks,there is often only limited face data,and an effective deep face model cannot be trained.Transfer learning is a process that uses the similarity between data and models to apply the models learned or trained in the source domain to the target domain,using large samples to solve the problem of small samples.Therefore,by migrating a pre-trained deep face model to a specific face recognition task,an effective deep face model can be trained using limited face data.Face recognition based on transfer learning has performed very well on the face data set under the condition of no offset on the front.However,because the facial images of some specific scenes are affected by changes in head state,angle,expression,and lighting,the recognition accuracy of these facial images still needs to be improved.First,for the max pooling layer and the average pooling layer,the contribution weight of each element of the face feature cannot be extracted in a balanced manner.On the basis of transfer learning,the paper uses the weighted average pooling layer to replace the last average pooling layer of the network model.By combining the output of the upper layer of the pooling layer with a learnable weight matrix,each element can transfer its own information after pooling,to ensure that the parameters with different contributions have different weight coefficients,which improves Face image recognition effect.Secondly,regarding the study of the face recognition loss function,in the face recognition task,different facial images of the same person are expected to be closer in the representation space,while the image features of different people are far apart.In view of the problem that the facial images of different identities are close to each other in the representation space,the paper proposes a center loss plus a constraint to encourage facial images of different identities to be farther in the representation space.The experimental results and visualization results show that the new loss function designed has a certain improvement in the accuracy of face recognition compared to the deep face model trained with the basic classification loss function.Finally,two sets of experiments were conducted.The first group of experiments was used to verify the effectiveness of the network model in extracting facial features through the weighted average pooling method.The second set of experiments was used to verify the effectiveness of the weighted average pooling method combined with the central loss function and constraints to improve the accuracy of face recognition.