Face Verification Algorithm Based On MS-CNN And Joint Bayesian

In recent years,with deep convolutional neural network developing,the performance of face verification improves rapidly.In the structure of general deep convolutional neural network,the restriction of the size of input maps and the number of down-sample operation can lead to that the depth of convolutional network and the level of abstraction of feature vectors are insufficient.In this paper,we propose a new multi-step convolutional neural network structure to solve above problems.We design a network which includes multiple convolution operations with single pooling operation.The number of consecutive convolution operation increase with the number of network layers increasing.In this way,we can reduce the calculation effectively.Moreover,we design a combination of loss function identification and verification to be a new loss function.The new loss function make the feature vectors which are extracted by MS-CNN distinguishable.Compared to traditional features designed by researchers,the feature vectors can make intra-personal variations smaller and make inter-personal variations larger automatically.We use Joint Bayesian classifier to recognize the feature vectors extracted by MS-CNN.In order to make the parameters of Joint Bayesian classifier convergent,we normalize feature vectors.When Joint Bayesian classifier is trained by a large number of face images,we can use a corresponding probability measure which is better than cosine distance to compare the feature vectors.We evaluate our algorithm on FERET dataset.Our method achieves the verification rates of 100%,100%,98.9%,99.7%on four sub-datasets in FERET which are Fb,Fc,Dup1,Dup2 when false accept rate(FAR)is 0.001.