Facial Image Quality Assessment Via Deep Learning

In recent years,face recognition technology is widely used in various fields.Under the conditions with control,face recognition systems have reached a very high recognition rate.However,in the conditions without control,the face image quality is reduced,which results in a decline in face recognition system performance.In order to solve this problem,we need to evaluate face image quality in advance.In this paper,we focus on facial image quality assessment based on the performance of face recognition.We compare several common face quality assessment algorithms and find out that they are not useful for face recognition.So we propose a new face image quality assessment algorithm.Firstly,we propose a novel procedure to design a reasonable and comprehensive face image quality metric.We use the VGG Face network model to extract face feature,and calculate the difference between the feature of test face image and that of the reference face image with PLDA.The difference is regard as the face quality value of the test facial image.We find out that occlusion and pose play more important roles in facial image quality assessment.Our experiment proves that the face image quality values can not only assess the face image quality,can also connect the facial image quality with the performance of face recognition.It effectively solves the shortage problem of quality score marks for the training data.Secondly,we introduce a no-reference image quality assessment based on convolution network.Considering the influence of the human eye vision system,we combine convolution network and saliency detection to effectively forecast the quality of image quality scores.Our experiment proves that this algorithm can effectively predict the image quality score,but cannot be directly applied to face image quality evaluation.Finally,we propose a no-reference facial image quality assessment via deep learning.We train an end-to-end deep convolutional neural network to automatically predict this quality value of the input facial images.We adopt the VGG-16 network as our network prototype,and fine-tune the VGG network parameters.In this way,our end-to-end system directly outputs a single overall general metric.Our algorithm is able to distinguish the better facial image from the worse ones,which could help to improve the performance of the face recognition and has good robustness.