Compressed Face Image Restoration Method Based On Depth Prior Constraint

Compressed image restoration is widely used in transportation and storage fields as a technology to reduce the amount of data and retain important information of images.At the same time,there are many needs to mine more information from compressed images,which leads to the birth of the field of lossy compressed image restoration.Over the years,great progress has been made in the research of compressed image restoration,which can recover hidden details from compressed images,but most of the algorithms are implemented for natural images,and the effect of compressed face image restoration is not satisfactory.As a special image category,compressed face images have structural prior information that natural images do not have.Making full use of the prior information in images to design networks can effectively improve the quality of restoration.To solve the above problems,this paper studies the compressed face image restoration method based on depth prior constraint.(1)Using the prior distribution of face components to guide the restoration of depth features,a back-projection conditional network based on pyramid structure is proposed,which makes the face components get clearer restoration results.The restoration branch is responsible for the main restoration work of the network,and the pyramid structure network is responsible for the feature extraction of different sizes of the image,and the back projection mechanism is responsible for the restoration details correction.The prior branch extracts prior information from the face component distribution map to guide the generation of depth features in the back-projection module.Experimental results show that compared with other restoration algorithms,this method not only has a higher objective evaluation index but also has clearer features of facial features.(2)According to the characteristics of the human visual processing system,a back-projection generation countermeasure network based on multi attention mechanism is proposed.The residual module of multi attention mechanism is introduced into the network to learn the importance of features,to reduce the interference of redundant features on the restoration process.At the same time,considering the poor visual perception effect of the restoration results optimized by MSE,the texture and details that conform to the human eye perception are generated by using the generation confrontation network structure.By introducing an attention mechanism in the shadow unit and feature fusion mechanism between subunits,the feature fusion and trade-off are optimized.The experimental results show that the back-projection generation countermeasure network based on multi attention mechanism can generate the restored image with finely and true details.(3)A dual network for blind restoration based on a high-quality face prior dictionary is proposed for blind restoration scenes where a large number of high-quality to low-quality image pairs cannot be obtained.In the training phase,first of all,we need to build a high-quality face component prior dictionary based on the existing high-quality face image database,combined with the facial features location information of the face image,to form a high-quality prior dictionary of the blind restoration network.Then a dual regression strategy is proposed to learn the reverse compression mapping from low quality to high quality,and the global back-projection mechanism is used to realize the training adjustment in blind restoration.The experimental results show that the dual network based on a highquality face prior dictionary can effectively use the high-quality prior information,and can also generate clear restored images without compressed information.To sum up,this paper studies and implements the restoration algorithm for different depths of prior constraints.For face images with training image pairs,it mainly considers the structural information such as facial features distribution,supplemented by back-projection mechanism,attention mechanism,and generating confrontation mechanism.For images without training image pairs,it uses the prior features of high-quality face components,Experiments show that this kind of deep prior constraint can bring better optimization results for the restoration algorithm.