Biologically Inspired Facial Image Quality Assessment And Enhancement

Facial images are of great significance for public security,such as identity recognition and case solving.However,due to the effect of the environmental lighting,system performance and human factors,the facial images captured in real world are always suffered from quality degradation problem,such as abnormal illuminations and low resolution.Such degradations brought a great challenge to the identification and recognition algorithm.Existing quality assessment research mainly focus on the perceived quality of natural images(such as clarity,aesthetics,etc.).However,facial images of actual scenes are often obtained under uncontrollable conditions(various types of noise,pose variation,etc.),and the biological quality of the image like the effectiveness for identity recognition is highly valued in the recognition task.At the same time,the most facial images captured in actual scenes are in different postures.What's worse,in the field of face synthesis,the traditional methods have various problems(image blurring,loss of identity information,loss of texture,etc.),especially the face images with extreme pose and complex expressions.These images often lose a lot of information,which are very important for identification and recognition.Therefore,this article will discuss two novel algorithms proposed through deep learning and generative adversarial network :(1)Biological quality assessment model of human face image based on the light convolution neural network.In order to evaluate the collected data properly,we used a deep learning method to train a robust distortion classifier based on the light convolution neural network with the Max-Feature-Map(MFM)activation layer.Further,the final quality assessment score was calculated according to the confidence of the image distortion classification and the face image recognition accuracy under the corresponding distortion.This method has the advantages of both the quality prediction and the biological quality.The obtained quality score has a positive relationship with the biological quality,which can well indicate the contribution of the image in the recognition system.Then,based on the predicted score,the appropriate images are selected from the large number of acquired images to identify and improve the effectiveness of the recognition system.(2)Face frontalization basing on the generative adversarial network.In order to reduce the impact of extreme postures on recognition performance and to retain enough original information to facilitate human reviewing,we preserved the light and expression in the image while using the proposed algorithm to get the frontalization facial image.In the algorithm design,we used the stack structure in generate adversarial network,and the profile images are synthesized and repainted to obtain the reasonable front view images.These images retain the identity information consistent with the original profile images and have greater contribution to the task of face recognition.