Research On Compressed Face Hallucination Based On Deep Convolutional Network

Face image is one of the most common natural images,and plays an important role in social media,video surveillance and other fields.However,in practical applications,the images in the social media and surveillance videos will be compressed,and the quality of the face image will be degraded by the influence of the compressed noise.In addition,in order to make the face details more clear to improve the accuracy of subsequent recognition,analysis and other tasks,it is usually necessary to superresolve the face images.In this paper,we study the super-resolution algorithm of compressed face images,considering joint denoising and super-resolution reconstruction.The super-resolution of compressed face images is a ill-posed problem,which requires additional prior knowledge.Face image has a strong domain knowledge.For the restoration of the degraded face images,we propose a data-driven method for mining the special structure information from a large number of face training images.At the same time,we use the strong prior constraints provided by the quantization bin.Accordingly,two deep convolutional network models for compressed face superresolution are proposed in this paper,specifically:we firstly propose a novel end to end convolutional neural network for super resolution of compressed face.We model the degradation process of compressed face images and propose a novel loss functions based on the degradation model.A good result is achieved by using the loss function.Secondly,according to the prior knowledge of the face,two network models with different architecture are designed: Bi-network and cascade network.The main idea is to detect the high frequency region of the face by using face landmarks.The high-frequency region is treated with special treatment.In the Bi-network,one of the branches is a common branch,which is used to reconstruct the smooth region of the face.The second branch is a high-frequency branch,which is used to restore the high-frequency region through the guidance of the high frequency prior of the face.In the cascade network,the model is divided into two stages to train.The first stage is to view the face image as a natural image.However,because the reconstruction results in the high frequency region are often smooth and lack of necessary details,we additionally utilize the prior knowledge as input and only penalize over the high-frequency region to recover the high frequency region,and then some of the texture details are synthesized.In addition,we also try to add perceptual loss to the loss function when training,and we will display these result.Extensive experiments demonstrate that the proposed algorithm performs favorably against the state-of-the-arthallucination method by a large margin on JPEG compressed face images,and among all methods,the proposed method achieves best testing results on the common benchmark dataset.