Research On Deep Learning Based Face Verification

Face verification is one of the most popular topics in computer vision and pattern recognition community.It is widely used for identity authentication in enormous areas such as finance,military,public security and so on.With the rapid development of deep learning and the growing data volume,the face verification ability of computers has al-ready surpassed the human beings on several benchmarks.As the performance of face verification models is almost saturated,how to further improve the performance becomes a very challenging problem.Since the loss function plays a key role in guiding the train-ing of deep neural networks,this dissertation delves into the loss functions,modifying the most widely used classification and metric learning loss functions to make them more suitable for face verification task.The main contribution of this dissertation includes the following aspects.1.Proposing to apply L₂hypersphere embedding in neural networks.The similarity used by the softmax cross-entropy loss is inner-product similarity.But during the deployment,the most common similarity measurement is cosine similarity.The difference between inner-product and cosine similarity is the vector L₂normalization.To make the training procedure consisting with the testing procedure,the inner-product layer should be replaced by the cosine layer.However,such a trivial modification would make the model unable to converge.This dissertation provides a theory to explain this problem.The theory describes the lower bound of softmax loss with regard to the norm of features.Since the norms of the features are normalized to 1,the lower bound is too high to make the model converge.According to this theory,the solution can be inferred.By appending a scale factor after the cosine layer,the model is able to converge.This dissertation also analyzes several properties of the scale factor,which would help understanding it.By finetuning the face verification for a few epochs using the modified loss function,the performance improves significantly on several benchmarks.2.Reformulating metric learning loss functions.When the traditional metric learning methods are introduced into deep learning,sam-pling is a crucial step to find hard pairs or triplets of samples to feed into the neural net-work.However,the sampling algorithms are usually very tricky and difficult to imple-ment.Based on the vector normalization,the weight matrix used by softmax cross-entropy loss is actually playing a role of“class agent”.By applying the class agent strategy on met-ric learning loss functions,the sampling problem can be avoided,while the advancements of metric learning loss functions are still retained.By analyzing the normalization operation in neural networks,it can be discovered that the feature normalization has an effect of hard sample mining and the weight normalization has an effect of alleviating the class imbalance problem.All the loss functions proposed in this dissertation have feature and weight normalization,so they all have such properties.Experiments show that the reformulated loss functions have superior performance than the traditional metric learning loss functions.3.Introducing the additive margin scheme into softmax cross-entropy loss.Based on the softmax cross-entropy loss with L₂hypersphere embedding,this disser-tation analyzes the similarities and differences between classification loss functions and metric learning loss functions.By introducing the margin strategy into softmax cross-entropy loss,the advancements of both classification and metric learning loss functions are merged into the proposed loss function.Among several margin schemes,additive margin achieves the best performance.This dissertation also proposes an evaluation metric named“latent margin”.With the latent margin,researchers will get a direct feeling about the discriminative power of the training model.The relationship between the latent margin and our proposed additive margin is also analyzed to guide the parameter tuning during the model training.By importing the additive margin,the performance of face verification model im-proves significantly on several benchmarks.It is the state-of-the-art face verification loss function by the time this dissertation is written.